Governance & Provenance

An agent you can audit is an agent you can trust.

Before an agent acts on your platform, every tool it can reach has been explicitly approved: pinned identity, typed capabilities, secrets it never sees, egress it can't exceed. While it acts, everything it does is written to an append-only provenance ledger. Capability is granted; nothing is assumed.

sentinel · tool policies
6Approved tools
denyDefault posture
0Config drift alerts

Policy · observability-readonly

Applied to agent · incident-triage

Active
Tool identitygrafana · pinned & approvedVerified
CapabilitiesQuery only · no writesTyped
SecretsHeld by the runtime, never the agentIsolated
EgressAllowlisted endpoints onlyEnforced

What the policy gates

Default-deny, enforced before the agent acts

Tool identity

Only approved tool servers, with pinned identity and validated configuration, exist for the agent. Drift from the approved state is caught before execution.

Typed capabilities

Capabilities are enumerable, not implied: which queries, which namespaces, read or write. A capability that wasn't granted can't be exercised.

Secrets, isolated

Credentials for your observability stack and clusters are held and injected by the runtime. They never pass through the agent's context.

Audit obligations

The policy states what must be logged, not just what is allowed. An action that can't meet its audit obligation doesn't run.

Approve, enforce, record

Governance as the execution path, not a wrapper

Provenance isn't a logging afterthought bolted onto an agent framework. In Sentinel, the governed path is the only path: approval happens before execution, enforcement happens at execution, and the record is produced by the runtime itself.

01 · APPROVE

Tools enter a registry before agents can touch them

A tool joins the platform the way a dependency should: reviewed, identified, and projected into a typed interface. Until that happens, no agent can call it, there is no ambient access to discover.

  • Approved server identity, pinned configuration
  • Typed projection of what the tool exposes
  • Default-deny: unregistered means unreachable
policy.ts
// capability is granted, never assumed
const policy = await sentinel.policies.create({
  name:  "observability-readonly",
  tools: ["grafana", "kubernetes"],
  caps:  ["query", "list"],   // no writes
  audit: "every_call"
})
→ default: deny · exceptions: enumerated
02 · ENFORCE

Every call is resolved at execution time

When the agent asks for a tool, the runtime resolves the request against the registry and the active policy, then executes it on the agent's behalf. An out-of-policy request never runs, and the refusal itself becomes part of the record.

  • Runtime-mediated execution, not agent-held credentials
  • Out-of-policy calls fail closed
  • Refusals are recorded alongside actions
enforcement · live
grafana.query · in policy executed
k8s.list pods · in policy executed
k8s.delete pod · not granted denied · logged
03 · RECORD

An append-only ledger of everything that happened

Tool calls and their outputs, model prompts and responses, evidence retrieved, configuration overlays, human interventions, each is an entry on a provenance ledger that only grows. From it you can reconstruct any investigation as one causal story: what triggered it, what the agent did, and how it reached its recommendation.

  • Append-only: no silent edits, ever
  • One causal story from trigger to recommendation
  • The raw material for audit, and for evaluation
provenance · ledger

INC-4212 · full trail

Trigger → evidence → report → review

Complete
Tool calls18 · with outputsRecorded
Model calls12 · prompts & responsesRecorded
Human review1 · diagnosis confirmedRecorded

Evaluation, unlocked

The trail is how you know the agent works

What the ledger captures

4
Tool calls & outputs
Model prompts & responses
Evidence artifacts
Human interventions

Judging an agent's work

trace
The answer alone not enough
The full execution trace evaluable

Edits to history

0
Silent edits possible 0
Append-only entries 100%

Give every agent action a paper trail.

Approve tools before agents can touch them, fail closed on everything else, and keep a record no one can quietly rewrite.