Product Brief · Runtime Governance for AI Agents

Sentience Governor

See what an agent was asked to do, what it actually did, and where the work moved outside scope.

Sentience Governor compares runtime actions with the objective and scope declared for the session and records the evidence.

Runs locally. Deterministic. Session-scoped. Non-blocking today.

Use cases

See How Governor Fits Real Agent Workflows

Start with the agent, what it is supposed to do, and the runtime actions your team or customers need to see.

Customer Support Agents

See when a support agent accesses or changes things outside the customer request.

Explore use case

Coding Agents

See when a coding agent touches files, tools, or systems outside the task.

Explore use case

Agentic Product Trust Screen

Give customers a simple Trust Screen showing what the agent did, what went as expected, and what Governor flagged.

Explore use case

Multi-Agent Runtime Dashboard

Govern multiple agents from one dashboard, built on Governor or delivered by Sentience.

Explore use case

Architecture

Put Governance Where Agent Actions Execute

Sentience Governor sits between the agent orchestrator and the model, tools, or optional gateway. It compares runtime actions with the session’s declared objective and scope, then records governance evidence without blocking execution.

Agent
Orchestrator
Sentience Governor
Execution Boundary
Evaluates · Records · Traces
Model / LLM
Gateway (optional)
Tools / APIs
Agent Execution Record
Structured governance trace written after each evaluated session
Policy Fingerprint
Deterministic reference linking execution to declared policy
Evidence Chain
Scope assertions, context snapshots, and governance results

Turn a Generic Trace into Evidence You Can Govern

Generic Trace
Timestamp
2025-08-21T14:23:07Z
Model
claude-3-5-sonnet
Input tokens
1842
Output tokens
312
Latency
1.4s
Status
success

Records what the model did. Does not capture intent, scope, policy, or governance context.

Agent Execution RecordSentience terminology
Declared intent
Refactor authentication module
Declared scope
src/auth/** (read/write)
Context snapshot
3 files examined, 1 write attempted
Memory write attempt
src/auth/session.ts ← flagged
Policy fingerprint
pol_v1_sha256_a3f9c1…
Governance result
RECORDED — scope match, intent declared
Evidence chain
Linked to policy fingerprint

Captures declared intent, scope assertions, context snapshots, memory write attempts, policy fingerprint, governance result, and evidence chain.

Evaluation scope

Connect Agent Actions to the Task and Scope They Started With

Guardrails shape model output. Observability records system activity. Sentience Governor connects an agent’s runtime actions to the objective and scope declared for that session.

Guardrails

Shape or evaluate model output.

Observability

Record system and application activity.

Runtime Governance

Compare agent actions with the objective and scope declared for the session, then record the result.

Example session — Customer A
Declared objective
"Refactor the authentication module to use the new session token format."
Declared scope
src/auth/** (read/write)
In scope — evaluated and recorded
  • Read src/auth/session.ts — within declared scope
  • Read src/auth/tokens.ts — within declared scope
  • Write src/auth/session.ts — within declared scope, recorded as memory write attempt
  • Context snapshot taken — 3 files examined
Outside scope — flagged and recorded
  • Read src/payments/stripe.ts — outside declared scope, flagged
  • Attempted access to environment variable DATABASE_URL — outside declared scope, flagged

Next step

Bring Us an Agent Workflow You Want to Govern

Show us what the agent is supposed to do, what it can act on, and where you need more visibility or control.

Talk to us →

Prefer to evaluate locally? Install from PyPI or read the docs.