Zephr

Governance / Agents

Agents that propose,humans that confirm.

Background and registered agents run under Guard: a hard deny-list at the scope boundary, a build-gate before any propose, evidence-gated writes, and a propose-via-PR default. The agent that proposes a change cannot confirm it — the fix lands as a human-reviewed pull request, never direct to main.

Evolving · built, hardeninghard deny-listbuild-gatepropose-via-PRnever self-confirm32 MCP tools · 11 platform surfaces · shared registry
Governed execution

Four gates. The sequence stops.

An autonomous action is not a weighted score across checks. It is a fixed sequence. Fail one gate and the rest are not reached. The diagram below fails at the evidence gate on purpose: that is the honest picture of a proposal that has not yet been anchored.

An autonomous action gated by a hard deny-listautonomous write enters at the left and is checked in a fixed order: deny-list, then build-gate, then evidence, then confirm. It stops at the first failing check at a hard boundary and carries away a typed reason. The remaining gate is planned but never reached. The sequence short-circuits at the hard boundary. A failed evidence gate never reaches human confirmation — and the proposing agent is never the confirming principal.hard stopactdeny-listpassedbuild-gatepassedevidencefailedconfirm—not reached
  1. 01deny-list

    Is this action on the hard refuse list?

    passed

  2. 02build-gate

    Did the change build and typecheck before propose?

    passed

  3. 03evidence

    Is the write anchored to evidence the human can open?

    failed

  4. 04confirm

    Is a different principal confirming the proposal?

    not reached

evidence_missing: write is not anchored to an openable evidence node

The sequence short-circuits at the hard boundary. A failed evidence gate never reaches human confirmation — and the proposing agent is never the confirming principal.

The four-gate sequence shown is illustrative — the real gates run in code.

Deny-list

Hard refusals, not advisory warnings.

Spinspire taught the industry that advisory deny lists fail under pressure. These patterns are refused at the scope boundary before any tool runs. The list is illustrative of the class, not an exhaustive public dump of every matcher.

Hard deny-list classes for autonomous agents
Pattern classWhy it is hard-refused
git push --force / force-with-lease to protected branchesRewrites shared history. Always refused at the scope boundary.
direct commits or merges to mainBypasses human review. Propose-via-PR is the only write path.
rsync / bulk filesystem overwrite of workspace rootsSilent mass mutation. No evidence trail the human can open.
destructive deletes of production data or secretsIrreversible. Agents may propose a plan; they may not execute it.
Surfaces

What is built, and what is still hardening.

Agent platform, A2A registration, and autonomous-agent runtime exist as code. Maturity is evolving until soak and operator runbooks catch up. The never-self-confirm rule is already a shipped constitutional constraint.

Agent product surfaces and maturity
SurfaceStatusTruth
Agent platformEvolvingRegistration, task ledger, and tool surface for user-facing agents.
A2A connectorEvolvingAgent-to-agent registration and card exchange under the same scope rules.
Autonomous agentsEvolvingGuard-governed background agents: deny-list, build-gate, evidence, PR-only writes.
Session-as-memoryShipped ruleAgent session memory is never self-confirmed. Confirmation is a human act.
Confirmation law

The proposer is never the confirmer.

This is the same two-lane confirmation law Trust and Review already state. Agents do not get a third lane.

  • Propose-via-PR default

    Autonomous writes land as pull requests. Direct-to-main is on the hard deny-list.

  • Evidence-gated writes

    A write without an openable evidence anchor fails the gate before a human is asked.

  • Never self-confirm

    Agent session memory is provisional until a human confirms. Constitution constraint 1.

  • Build-gate first

    A change that does not build or typecheck is not proposed. The gate is mechanical.

Create, run, list and inspect autonomous agents

zephr agent

Two lanes, one direction

  1. proposer — the agentproposes via pull request
  2. gate — build + evidencemust pass before review
  3. confirmer — a humanthe only principal that confirms

The proposing agent is never the confirming principal. Direct-to-main is on the hard deny-list; the only write path is a pull request a person reviews.

Illustrative example — not live data

See Guard, or read how trust is measured.

Agents inherit Guard's scope and Review's confirmation law. The trust page states what is measured today and what is still not evaluated.