Question 1

What does Sentience do?

Accepted Answer

Sentience evaluates what an AI agent actually does against what it said it would do. Before acting, an agent declares an objective and the scope it needs. Sentience then captures structured events at the execution boundary (registration, intent, scope, context, memory writes) and evaluates each action against that declaration at the moment it happens.

Question 2

How is Sentience different from logs or observability tools?

Accepted Answer

Traditional logs and observability primarily record what happened. Sentience also captures the agent's declared intent, asserted scope and runtime context, then evaluates actions against that declaration during execution. The result is a structured record that humans and policy systems can reason about, rather than a stream of events to search through afterwards.

Question 3

Does Sentience require an API, account, or external service?

Accepted Answer

No. Sentience runs entirely locally as a Python package. It does not require an API, account creation, or an external service. By default nothing leaves your machine. An operator can configure a sink to send events somewhere else, but that only happens if you set it up.

Question 4

What data does Sentience inspect?

Accepted Answer

Sentience reads the agent's declared intent and scope, and for each action the tool name, target system, operation type and classification metadata. It does not inspect raw payload content unless the integrator explicitly provides a classification hook.

Question 5

Does Sentience block or control agent actions?

Accepted Answer

No. Sentience records and evaluates. It does not block, pause or modify agent actions. When an action falls outside the declared scope, Sentience raises a policy violation in the trace and the action still proceeds. Enforcement is part of Sentience's product direction and is being developed with public input.

Question 6

What are the core signals Sentience captures?

Accepted Answer

Five control points produce structured governance events: agent registration, intent declaration, scope assertion, context snapshot and memory write attempt. Events also carry execution telemetry for the turn they belong to: token counts, model and provider. That lets teams associate compute spend with the turn and action context where it occurred, rather than only with a session total.

Question 7

How flexible is the governance policy?

Accepted Answer

Policy is an operator-authored YAML profile. It controls when intent must be declared, which task-boundary signals are active (directory change, file-type shift, time gap, read-to-write transition), and which tool patterns count as high-consequence. Today each machine uses one configured profile. It is bound into each session and its fingerprint is recorded with the resulting evidence, so Sentience can show which policy applied to which run. Every match currently produces a flag. Scoping policy per team or environment, and acting differently by risk tier, is the enterprise direction.

Question 8

What does the output look like?

Accepted Answer

An append-only JSONL trace of governance events. Each event is sequence-chained to the one before it, so gaps and reordering are detectable. Policy violations and advisory flags are reported separately from one another, and the sentience CLI renders a session view for quick inspection.

Question 9

How do teams integrate Sentience?

Accepted Answer

Sentience integrates with existing agent setups without requiring rewrites: Claude Code hooks and slash commands, an MCP client wrapper, an opt-in MCP server, and a LangChain callback handler with LangGraph support.

Question 10

Is Sentience replacing model safety or application logging?

Accepted Answer

No. They answer different questions. Model safety governs model inputs and outputs. Observability records execution. Sentience evaluates runtime actions against the governance context declared for the session.

Question 11

Who is Sentience for?

Accepted Answer

Developers who want to understand agent behaviour, operators who need visibility across systems, and enterprises preparing for governance and enforcement.

Question 12

What is Sentience's direction beyond visibility?

Accepted Answer

The same execution-boundary control points used for tracing and evaluation are designed to support enforcement, operator controls and action-level policy.

Frequently asked questions