Production tracing for agents: observability without a data exhaust problem addresses one concrete problem: Design traces that answer debugging questions while limiting sensitive model and tool payload retention. The guide works from the actual objects — trace, span, sensitive data, metadata, retention — and aims for a verifiable decision rather than a generic pattern.
The concrete problem: trace meets span
The hard part appears when the happy path meets authorization, failures, and operational constraints.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c11 starts at retention and treats trace as an explicit boundary rather than an implicit assumption. The first control requires span to emit an observable result before sensitive data can trigger the intended effect around metadata. Operations then observes the transition between span and sensitive data, while security checks that metadata receives neither implicit authority nor unnecessary data. If the sensitive data verification fails, rollback restores the configuration around metadata, replays agent-production-tracing-data-minimization-c11, and compares the new state with the control evidence from retention. This detail makes “Production tracing for agents: observability without a data exhaust problem” reviewable because each operational statement points to trace, a concrete condition, and evidence instead of a generic assurance.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c12 starts at trace and treats span as an explicit boundary rather than an implicit assumption. The first control requires sensitive data to emit an observable result before metadata can trigger the intended effect around retention. The conclusion stays bounded by sensitive data and retention: anything not demonstrated by scenario agent-production-tracing-data-minimization-c12 is labeled as a limitation or inference, never promoted to fact. That sequence serves this concrete job: Design traces that answer debugging questions while limiting sensitive model and tool payload retention. It turns trace into a reviewable decision point with a named input and a retained output. To test span, fixture agent-production-tracing-data-minimization-c12 contains both an allowed state and a rejected state; rejection must occur before any change attributed to sensitive data.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c13 starts at span and treats sensitive data as an explicit boundary rather than an implicit assumption. The first control requires metadata to emit an observable result before retention can trigger the intended effect around trace. Operations then observes the transition between metadata and retention, while security checks that trace receives neither implicit authority nor unnecessary data. If the retention verification fails, rollback restores the configuration around trace, replays agent-production-tracing-data-minimization-c13, and compares the new state with the control evidence from span. This detail makes “Production tracing for agents: observability without a data exhaust problem” reviewable because each operational statement points to sensitive data, a concrete condition, and evidence instead of a generic assurance.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c14 starts at sensitive data and treats metadata as an explicit boundary rather than an implicit assumption. The first control requires retention to emit an observable result before trace can trigger the intended effect around span. The conclusion stays bounded by retention and span: anything not demonstrated by scenario agent-production-tracing-data-minimization-c14 is labeled as a limitation or inference, never promoted to fact. That sequence serves this concrete job: Design traces that answer debugging questions while limiting sensitive model and tool payload retention. It turns sensitive data into a reviewable decision point with a named input and a retained output. To test metadata, fixture agent-production-tracing-data-minimization-c14 contains both an allowed state and a rejected state; rejection must occur before any change attributed to retention.
Trust boundaries around sensitive data
A robust implementation separates what the model proposes from what the application authorizes and verifies.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c21 starts at trace and treats span as an explicit boundary rather than an implicit assumption. The first control requires sensitive data to emit an observable result before metadata can trigger the intended effect around retention. To test span, fixture agent-production-tracing-data-minimization-c21 contains both an allowed state and a rejected state; rejection must occur before any change attributed to sensitive data. Operations then observes the transition between sensitive data and metadata, while security checks that retention receives neither implicit authority nor unnecessary data. If the metadata verification fails, rollback restores the configuration around retention, replays agent-production-tracing-data-minimization-c21, and compares the new state with the control evidence from trace.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c22 starts at span and treats sensitive data as an explicit boundary rather than an implicit assumption. The first control requires metadata to emit an observable result before retention can trigger the intended effect around trace. This detail makes “Production tracing for agents: observability without a data exhaust problem” reviewable because each operational statement points to sensitive data, a concrete condition, and evidence instead of a generic assurance. The conclusion stays bounded by metadata and trace: anything not demonstrated by scenario agent-production-tracing-data-minimization-c22 is labeled as a limitation or inference, never promoted to fact. That sequence serves this concrete job: Design traces that answer debugging questions while limiting sensitive model and tool payload retention. It turns span into a reviewable decision point with a named input and a retained output.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c23 starts at sensitive data and treats metadata as an explicit boundary rather than an implicit assumption. The first control requires retention to emit an observable result before trace can trigger the intended effect around span. To test metadata, fixture agent-production-tracing-data-minimization-c23 contains both an allowed state and a rejected state; rejection must occur before any change attributed to retention. Operations then observes the transition between retention and trace, while security checks that span receives neither implicit authority nor unnecessary data. If the trace verification fails, rollback restores the configuration around span, replays agent-production-tracing-data-minimization-c23, and compares the new state with the control evidence from sensitive data.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c24 starts at metadata and treats retention as an explicit boundary rather than an implicit assumption. The first control requires trace to emit an observable result before span can trigger the intended effect around sensitive data. This detail makes “Production tracing for agents: observability without a data exhaust problem” reviewable because each operational statement points to retention, a concrete condition, and evidence instead of a generic assurance. The conclusion stays bounded by trace and sensitive data: anything not demonstrated by scenario agent-production-tracing-data-minimization-c24 is labeled as a limitation or inference, never promoted to fact. That sequence serves this concrete job: Design traces that answer debugging questions while limiting sensitive model and tool payload retention. It turns metadata into a reviewable decision point with a named input and a retained output.
Build the decision path with metadata
The starting point is not a feature; it is an observable decision boundary.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c31 starts at span and treats sensitive data as an explicit boundary rather than an implicit assumption. The first control requires metadata to emit an observable result before retention can trigger the intended effect around trace. That sequence serves this concrete job: Design traces that answer debugging questions while limiting sensitive model and tool payload retention. It turns span into a reviewable decision point with a named input and a retained output. To test sensitive data, fixture agent-production-tracing-data-minimization-c31 contains both an allowed state and a rejected state; rejection must occur before any change attributed to metadata. Operations then observes the transition between metadata and retention, while security checks that trace receives neither implicit authority nor unnecessary data.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c32 starts at sensitive data and treats metadata as an explicit boundary rather than an implicit assumption. The first control requires retention to emit an observable result before trace can trigger the intended effect around span. If the trace verification fails, rollback restores the configuration around span, replays agent-production-tracing-data-minimization-c32, and compares the new state with the control evidence from sensitive data. This detail makes “Production tracing for agents: observability without a data exhaust problem” reviewable because each operational statement points to metadata, a concrete condition, and evidence instead of a generic assurance. The conclusion stays bounded by retention and span: anything not demonstrated by scenario agent-production-tracing-data-minimization-c32 is labeled as a limitation or inference, never promoted to fact.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c33 starts at metadata and treats retention as an explicit boundary rather than an implicit assumption. The first control requires trace to emit an observable result before span can trigger the intended effect around sensitive data. That sequence serves this concrete job: Design traces that answer debugging questions while limiting sensitive model and tool payload retention. It turns metadata into a reviewable decision point with a named input and a retained output. To test retention, fixture agent-production-tracing-data-minimization-c33 contains both an allowed state and a rejected state; rejection must occur before any change attributed to trace. Operations then observes the transition between trace and span, while security checks that sensitive data receives neither implicit authority nor unnecessary data.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c34 starts at retention and treats trace as an explicit boundary rather than an implicit assumption. The first control requires span to emit an observable result before sensitive data can trigger the intended effect around metadata. If the sensitive data verification fails, rollback restores the configuration around metadata, replays agent-production-tracing-data-minimization-c34, and compares the new state with the control evidence from retention. This detail makes “Production tracing for agents: observability without a data exhaust problem” reviewable because each operational statement points to trace, a concrete condition, and evidence instead of a generic assurance. The conclusion stays bounded by span and metadata: anything not demonstrated by scenario agent-production-tracing-data-minimization-c34 is labeled as a limitation or inference, never promoted to fact.
Verify retention with observable evidence
A production design should make it clear who decides, what evidence is available, and what can be rolled back.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c41 starts at sensitive data and treats metadata as an explicit boundary rather than an implicit assumption. The first control requires retention to emit an observable result before trace can trigger the intended effect around span. The conclusion stays bounded by retention and span: anything not demonstrated by scenario agent-production-tracing-data-minimization-c41 is labeled as a limitation or inference, never promoted to fact. That sequence serves this concrete job: Design traces that answer debugging questions while limiting sensitive model and tool payload retention. It turns sensitive data into a reviewable decision point with a named input and a retained output. To test metadata, fixture agent-production-tracing-data-minimization-c41 contains both an allowed state and a rejected state; rejection must occur before any change attributed to retention.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c42 starts at metadata and treats retention as an explicit boundary rather than an implicit assumption. The first control requires trace to emit an observable result before span can trigger the intended effect around sensitive data. Operations then observes the transition between trace and span, while security checks that sensitive data receives neither implicit authority nor unnecessary data. If the span verification fails, rollback restores the configuration around sensitive data, replays agent-production-tracing-data-minimization-c42, and compares the new state with the control evidence from metadata. This detail makes “Production tracing for agents: observability without a data exhaust problem” reviewable because each operational statement points to retention, a concrete condition, and evidence instead of a generic assurance.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c43 starts at retention and treats trace as an explicit boundary rather than an implicit assumption. The first control requires span to emit an observable result before sensitive data can trigger the intended effect around metadata. The conclusion stays bounded by span and metadata: anything not demonstrated by scenario agent-production-tracing-data-minimization-c43 is labeled as a limitation or inference, never promoted to fact. That sequence serves this concrete job: Design traces that answer debugging questions while limiting sensitive model and tool payload retention. It turns retention into a reviewable decision point with a named input and a retained output. To test trace, fixture agent-production-tracing-data-minimization-c43 contains both an allowed state and a rejected state; rejection must occur before any change attributed to span.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c44 starts at trace and treats span as an explicit boundary rather than an implicit assumption. The first control requires sensitive data to emit an observable result before metadata can trigger the intended effect around retention. Operations then observes the transition between sensitive data and metadata, while security checks that retention receives neither implicit authority nor unnecessary data. If the metadata verification fails, rollback restores the configuration around retention, replays agent-production-tracing-data-minimization-c44, and compares the new state with the control evidence from trace. This detail makes “Production tracing for agents: observability without a data exhaust problem” reviewable because each operational statement points to span, a concrete condition, and evidence instead of a generic assurance.
Failure modes, signals, and diagnosis
The hard part appears when the happy path meets authorization, failures, and operational constraints.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c51 starts at metadata and treats retention as an explicit boundary rather than an implicit assumption. The first control requires trace to emit an observable result before span can trigger the intended effect around sensitive data. This detail makes “Production tracing for agents: observability without a data exhaust problem” reviewable because each operational statement points to retention, a concrete condition, and evidence instead of a generic assurance. The conclusion stays bounded by trace and sensitive data: anything not demonstrated by scenario agent-production-tracing-data-minimization-c51 is labeled as a limitation or inference, never promoted to fact. That sequence serves this concrete job: Design traces that answer debugging questions while limiting sensitive model and tool payload retention. It turns metadata into a reviewable decision point with a named input and a retained output.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c52 starts at retention and treats trace as an explicit boundary rather than an implicit assumption. The first control requires span to emit an observable result before sensitive data can trigger the intended effect around metadata. To test trace, fixture agent-production-tracing-data-minimization-c52 contains both an allowed state and a rejected state; rejection must occur before any change attributed to span. Operations then observes the transition between span and sensitive data, while security checks that metadata receives neither implicit authority nor unnecessary data. If the sensitive data verification fails, rollback restores the configuration around metadata, replays agent-production-tracing-data-minimization-c52, and compares the new state with the control evidence from retention.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c53 starts at trace and treats span as an explicit boundary rather than an implicit assumption. The first control requires sensitive data to emit an observable result before metadata can trigger the intended effect around retention. This detail makes “Production tracing for agents: observability without a data exhaust problem” reviewable because each operational statement points to span, a concrete condition, and evidence instead of a generic assurance. The conclusion stays bounded by sensitive data and retention: anything not demonstrated by scenario agent-production-tracing-data-minimization-c53 is labeled as a limitation or inference, never promoted to fact. That sequence serves this concrete job: Design traces that answer debugging questions while limiting sensitive model and tool payload retention. It turns trace into a reviewable decision point with a named input and a retained output.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c54 starts at span and treats sensitive data as an explicit boundary rather than an implicit assumption. The first control requires metadata to emit an observable result before retention can trigger the intended effect around trace. To test sensitive data, fixture agent-production-tracing-data-minimization-c54 contains both an allowed state and a rejected state; rejection must occur before any change attributed to metadata. Operations then observes the transition between metadata and retention, while security checks that trace receives neither implicit authority nor unnecessary data. If the retention verification fails, rollback restores the configuration around trace, replays agent-production-tracing-data-minimization-c54, and compares the new state with the control evidence from span.
Progressive rollout and rollback
A robust implementation separates what the model proposes from what the application authorizes and verifies.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c61 starts at retention and treats trace as an explicit boundary rather than an implicit assumption. The first control requires span to emit an observable result before sensitive data can trigger the intended effect around metadata. If the sensitive data verification fails, rollback restores the configuration around metadata, replays agent-production-tracing-data-minimization-c61, and compares the new state with the control evidence from retention. This detail makes “Production tracing for agents: observability without a data exhaust problem” reviewable because each operational statement points to trace, a concrete condition, and evidence instead of a generic assurance. The conclusion stays bounded by span and metadata: anything not demonstrated by scenario agent-production-tracing-data-minimization-c61 is labeled as a limitation or inference, never promoted to fact.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c62 starts at trace and treats span as an explicit boundary rather than an implicit assumption. The first control requires sensitive data to emit an observable result before metadata can trigger the intended effect around retention. That sequence serves this concrete job: Design traces that answer debugging questions while limiting sensitive model and tool payload retention. It turns trace into a reviewable decision point with a named input and a retained output. To test span, fixture agent-production-tracing-data-minimization-c62 contains both an allowed state and a rejected state; rejection must occur before any change attributed to sensitive data. Operations then observes the transition between sensitive data and metadata, while security checks that retention receives neither implicit authority nor unnecessary data.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c63 starts at span and treats sensitive data as an explicit boundary rather than an implicit assumption. The first control requires metadata to emit an observable result before retention can trigger the intended effect around trace. If the retention verification fails, rollback restores the configuration around trace, replays agent-production-tracing-data-minimization-c63, and compares the new state with the control evidence from span. This detail makes “Production tracing for agents: observability without a data exhaust problem” reviewable because each operational statement points to sensitive data, a concrete condition, and evidence instead of a generic assurance. The conclusion stays bounded by metadata and trace: anything not demonstrated by scenario agent-production-tracing-data-minimization-c63 is labeled as a limitation or inference, never promoted to fact.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c64 starts at sensitive data and treats metadata as an explicit boundary rather than an implicit assumption. The first control requires retention to emit an observable result before trace can trigger the intended effect around span. That sequence serves this concrete job: Design traces that answer debugging questions while limiting sensitive model and tool payload retention. It turns sensitive data into a reviewable decision point with a named input and a retained output. To test metadata, fixture agent-production-tracing-data-minimization-c64 contains both an allowed state and a rejected state; rejection must occur before any change attributed to retention. Operations then observes the transition between retention and trace, while security checks that span receives neither implicit authority nor unnecessary data.
Production decision criteria
The starting point is not a feature; it is an observable decision boundary.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c71 starts at trace and treats span as an explicit boundary rather than an implicit assumption. The first control requires sensitive data to emit an observable result before metadata can trigger the intended effect around retention. Operations then observes the transition between sensitive data and metadata, while security checks that retention receives neither implicit authority nor unnecessary data. If the metadata verification fails, rollback restores the configuration around retention, replays agent-production-tracing-data-minimization-c71, and compares the new state with the control evidence from trace. This detail makes “Production tracing for agents: observability without a data exhaust problem” reviewable because each operational statement points to span, a concrete condition, and evidence instead of a generic assurance.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c72 starts at span and treats sensitive data as an explicit boundary rather than an implicit assumption. The first control requires metadata to emit an observable result before retention can trigger the intended effect around trace. The conclusion stays bounded by metadata and trace: anything not demonstrated by scenario agent-production-tracing-data-minimization-c72 is labeled as a limitation or inference, never promoted to fact. That sequence serves this concrete job: Design traces that answer debugging questions while limiting sensitive model and tool payload retention. It turns span into a reviewable decision point with a named input and a retained output. To test sensitive data, fixture agent-production-tracing-data-minimization-c72 contains both an allowed state and a rejected state; rejection must occur before any change attributed to metadata.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c73 starts at sensitive data and treats metadata as an explicit boundary rather than an implicit assumption. The first control requires retention to emit an observable result before trace can trigger the intended effect around span. Operations then observes the transition between retention and trace, while security checks that span receives neither implicit authority nor unnecessary data. If the trace verification fails, rollback restores the configuration around span, replays agent-production-tracing-data-minimization-c73, and compares the new state with the control evidence from sensitive data. This detail makes “Production tracing for agents: observability without a data exhaust problem” reviewable because each operational statement points to metadata, a concrete condition, and evidence instead of a generic assurance.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c74 starts at metadata and treats retention as an explicit boundary rather than an implicit assumption. The first control requires trace to emit an observable result before span can trigger the intended effect around sensitive data. The conclusion stays bounded by trace and sensitive data: anything not demonstrated by scenario agent-production-tracing-data-minimization-c74 is labeled as a limitation or inference, never promoted to fact. That sequence serves this concrete job: Design traces that answer debugging questions while limiting sensitive model and tool payload retention. It turns metadata into a reviewable decision point with a named input and a retained output. To test retention, fixture agent-production-tracing-data-minimization-c74 contains both an allowed state and a rejected state; rejection must occur before any change attributed to trace.
Evidence point: OpenAI documents that mCP integrations can expose remote or local tools; the documentation recommends trusted servers, least-privilege credentials and approvals for sensitive operations. This source anchors production decision criteria but does not replace local verification. [S7]Controls that remain after launch
A production design should make it clear who decides, what evidence is available, and what can be rolled back.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c81 starts at span and treats sensitive data as an explicit boundary rather than an implicit assumption. The first control requires metadata to emit an observable result before retention can trigger the intended effect around trace. To test sensitive data, fixture agent-production-tracing-data-minimization-c81 contains both an allowed state and a rejected state; rejection must occur before any change attributed to metadata. Operations then observes the transition between metadata and retention, while security checks that trace receives neither implicit authority nor unnecessary data. If the retention verification fails, rollback restores the configuration around trace, replays agent-production-tracing-data-minimization-c81, and compares the new state with the control evidence from span.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c82 starts at sensitive data and treats metadata as an explicit boundary rather than an implicit assumption. The first control requires retention to emit an observable result before trace can trigger the intended effect around span. This detail makes “Production tracing for agents: observability without a data exhaust problem” reviewable because each operational statement points to metadata, a concrete condition, and evidence instead of a generic assurance. The conclusion stays bounded by retention and span: anything not demonstrated by scenario agent-production-tracing-data-minimization-c82 is labeled as a limitation or inference, never promoted to fact. That sequence serves this concrete job: Design traces that answer debugging questions while limiting sensitive model and tool payload retention. It turns sensitive data into a reviewable decision point with a named input and a retained output.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c83 starts at metadata and treats retention as an explicit boundary rather than an implicit assumption. The first control requires trace to emit an observable result before span can trigger the intended effect around sensitive data. To test retention, fixture agent-production-tracing-data-minimization-c83 contains both an allowed state and a rejected state; rejection must occur before any change attributed to trace. Operations then observes the transition between trace and span, while security checks that sensitive data receives neither implicit authority nor unnecessary data. If the span verification fails, rollback restores the configuration around sensitive data, replays agent-production-tracing-data-minimization-c83, and compares the new state with the control evidence from metadata.
In “Production tracing for agents: observability without a data exhaust problem”, scenario agent-production-tracing-data-minimization-c84 starts at retention and treats trace as an explicit boundary rather than an implicit assumption. The first control requires span to emit an observable result before sensitive data can trigger the intended effect around metadata. This detail makes “Production tracing for agents: observability without a data exhaust problem” reviewable because each operational statement points to trace, a concrete condition, and evidence instead of a generic assurance. The conclusion stays bounded by span and metadata: anything not demonstrated by scenario agent-production-tracing-data-minimization-c84 is labeled as a limitation or inference, never promoted to fact. That sequence serves this concrete job: Design traces that answer debugging questions while limiting sensitive model and tool payload retention. It turns retention into a reviewable decision point with a named input and a retained output.
Evidence point: OpenAI documents that encryptedSession can wrap a session store with Fernet encryption, per-session HKDF-derived keys and TTL-based expiration. This source anchors controls that remain after launch but does not replace local verification. [S8]Operational checklist
- The control around trace has an input, a rule, a rejection behavior, and evidence.
- The control around span has an input, a rule, a rejection behavior, and evidence.
- The control around sensitive data has an input, a rule, a rejection behavior, and evidence.
- The control around metadata has an input, a rule, a rejection behavior, and evidence.
- The control around retention has an input, a rule, a rejection behavior, and evidence.
Sources and control points
- [S1] OpenAI Agents SDK — The Agents SDK is a higher-level runtime around model calls that manages tools, guardrails, handoffs, sessions and tracing. source
- [S2] Agents - OpenAI Agents SDK — An Agent combines instructions, tools and optional runtime behavior such as handoffs, guardrails and structured outputs. source
- [S3] Tools - OpenAI Agents SDK — The SDK distinguishes hosted tools, local execution tools, function tools and agents-as-tools, each with different trust and execution boundaries. source
- [S4] Handoffs - OpenAI Agents SDK — Handoffs transfer the active conversation to a specialist agent and can filter or reshape the history passed to the destination. source
- [S5] Agent orchestration - OpenAI Agents SDK — Two common orchestration patterns are manager-controlled agents-as-tools and handoffs where a specialist becomes the active agent. source
- [S6] Guardrails - OpenAI Agents SDK — Agent and tool guardrails can validate inputs or outputs and can stop execution with tripwire-style failures. source
- [S7] Model context protocol (MCP) - OpenAI Agents SDK — MCP integrations can expose remote or local tools; the documentation recommends trusted servers, least-privilege credentials and approvals for sensitive operations. source
- [S8] Encrypted session - OpenAI Agents SDK — EncryptedSession can wrap a session store with Fernet encryption, per-session HKDF-derived keys and TTL-based expiration. source
- [S9] Results - OpenAI Agents SDK — Run results expose final output, new items, agent identity, raw responses, guardrail results, state and usage diagnostics. source
- [S10] AI Risk Management Framework — NIST positions the AI RMF as a voluntary framework for managing AI risks and is revising it while adding profiles for specific settings. source
Comments
No published comment yet.
Sign in to comment