Expand description
Harness runtime: governed execution for effect-producing agents.
A harness sits outside an AgentKernel. The kernel
decides the next action; the harness runtime executes provider requests
and routes side-effect requests through policy, approval, executor,
output limiting, audit, and transcript recording before feeding
observations back to the kernel.
This module deliberately does not include production filesystem, shell, git, browser, or MCP executors. Those belong in higher layers such as a coding-workload SDK. The built-in executors here are for tests, applications that provide their own effect semantics, and validating the runtime boundary.
Modules§
- agent
- Agents: reasoning loops.
- effect
- Effect protocol: side-effect requests and observations.
- message
- Conversation messages.
- observability
- Lightweight observability record types.
- provider
- Provider: the interface for communicating with an LLM.
- run
- Structured run protocol shared by agent implementations and callers.
- tool
- Tool registry support for agent runtimes.
Structs§
- Always
Allow Approval Broker - Approval broker that always allows requests.
- Always
Deny Approval Broker - Approval broker that always denies requests.
- Approval
Request - Approval request passed to an
ApprovalBroker. - Basic
Harness - Minimal in-process harness implementation.
- Classified
Effect - Classified effect request.
- Default
Policy Engine - Risk-based default policy.
- Default
Risk Classifier - Conservative default risk classifier.
- Execution
Policy - Execution policy passed to an
EffectExecutor. - Execution
Policy Summary - Serializable summary of an execution policy.
- Harness
Config - Harness configuration.
- Harness
Runtime - Outer runtime that drives an
AgentKernelwith aProviderand governedHarness. - Harness
Runtime Config - Runtime loop configuration.
- Limited
Output - Output after limiting and redaction.
- Model
Summary - Summary of a model observation.
- Noop
Audit Sink - Explicit opt-out audit sink.
- Noop
Effect Executor - Executor that refuses every effect without performing side effects.
- Noop
Redactor - Redactor that leaves output unchanged.
- Noop
Transcript Store - Transcript store that drops all records.
- Output
Limit - Output size limits.
- Pattern
Redactor - Secret-pattern redactor for examples and tests.
- RawEffect
Output - Raw executor output before limiter/redactor processing.
- Redacted
Text - Redacted text and metadata.
- Redaction
Record - A redaction that was applied to an exported record or text field.
- Router
Effect Executor - Executor that dispatches by
EffectKind. - Static
Approval Broker - Static approval broker configured with a single decision.
- Static
Effect Executor - Test executor that returns preconfigured outputs by effect id.
- VecAudit
Sink - In-memory audit sink useful for tests.
- VecTranscript
Store - In-memory transcript store useful for tests.
Enums§
- Agent
Action Summary - Summary of an agent action for transcript records.
- Approval
Decision - Approval decision.
- Approval
Error - Approval errors.
- Audit
Error - Audit errors.
- Audit
Event - Reliable effect-governance audit event.
- Execution
Error - Executor errors.
- Harness
Error - Errors returned by a harness.
- Harness
Runtime Error - Errors returned by
HarnessRuntime. - Network
Policy - Network access policy requested of an executor.
- Policy
Decision - Policy decision for a classified effect.
- Sandbox
Policy - Filesystem/process sandbox policy requested of an executor.
- Transcript
Error - Transcript errors.
- Transcript
Record - Transcript record for run replay and debugging.
Traits§
- Approval
Broker - Broker that obtains approval from an application-specific authority.
- Audit
Sink - Reliable audit sink.
- Effect
Executor - Executes an already approved effect.
- Harness
- Governs and executes one or more effect requests.
- Policy
Engine - Evaluates host policy for a classified effect.
- Redactor
- Redacts executor output before model/audit/transcript use.
- Risk
Classifier - Classifies effect risk before policy evaluation.
- Transcript
Store - Transcript store for resumable run traces.