AI Agent Governance ยท The Agentic Supervisory Plane

Give AI agents authority. Keep control.

Signet is the local, deterministic gate that decides pass, block, or hold on every action your agents take, and seals each one to a permanent, tamper-evident record. Runs in microseconds, inside your perimeter.

No per-action meter. No cap. A record that never expires.

Live in production today. Validated on public benchmarks we did not write.

S Signet gate
Deciding
0
actions checked
0
blocked pre-impact
0
held for a human

Illustrative stream. In production every decision is sealed to a hash-chained ledger.

Live
in production at Lavish Gains, governing a real agent fleet
Near zero
injected-attack success on public benchmarks, at about 1x overhead
~2 µs
gate decision, deterministic, no model in the call
Permanent
every action sealed on a tamper-evident, hash-chained ledger

The problem

AI agents now act. Nothing is checking whether they should.

Agents are handed real authority: to send, to spend, to delete, to change systems. Between the model deciding and the action happening, most stacks have nothing. Prompt guards are advisory, logs are after the fact, and identity tools were built for people, not for software that reasons and acts on its own.

~/.signet/notebooks/fleet.ledger
AgentActionProvenanceVerdictSealed
dev-agent-07write_file(config.py)owner-authorizedPassa91f72…
server-apiwrite_file(key=sk-live…)secret in payloadBlock3d0c81…
support-agentsend_email(external)untrusted-derivedHold7f1a9c…
finance-agentwire_transfer($5,000)not in mandateBlockb7375f…
ops-agentget_calendar(week)in scopePass2d8a6e…

The real ledger. Every verdict carries its reason and a hash that chains to the one before it, so the record cannot be edited after the fact.

How Signet works

Onboard the agent. Set its authority. Gate every action.

Onboard

A mandate at birth

Every agent is issued a sealed mandate: its role, scope, and limits, fixed before it acts. An identity that survives models, sessions, and systems, and that the agent itself cannot edit.

Set authority

Deterministic, not a model

Admission by set membership, arguments by provenance. The gate makes its call with logic, not a second model, so there is nothing to talk out of it with a reworded prompt. Overhead is about 1x.

Execute safely

Pass, block, or hold

Every action is decided at the moment of execution and written to an append-only, hash-chained record anyone can verify. Legitimate work passes, out-of-mandate actions are blocked, consequential ones are held for a human.

No per-action meter. No cap. A record that never expires.
Metered cloud control planes bill every action and delete your audit trail when the plan lapses. Signet runs locally and prices by the seat, so a busy agent is not a bigger bill, and the proof is yours to keep.

Proof

Running in production. Validated by outsiders.

Signet is not a concept. It governs a live fleet in production today, and its safety has been measured on public benchmarks that TrajeXY did not write. Every figure is recomputed from raw output, not taken on trust.

20+
agents governed in production
43,179
sealed decisions on the ledger
~0%
injected-attack success, benchmarks
~2 µs
per gate decision

Autonomy you can trust

Put a gate between your agents and execution.

See Signet governing agents live, and what it would take to deploy inside your stack. We will get back to you within one business day.