Agent governance
Agent governance is the set of controls that decide what an agent is allowed to do, prove what it did, and make the result reversible. Everything below is implemented in public repositories on this domain, with the patterns and limits written down.
What governance means here
Most agent demos fail in production for the same reason: nothing between the operator and the model is deterministic. Governance is the layer that makes an agent run auditable and bounded before it is autonomous. In practice that is seven controls.
- A deterministic policy decision runs before any provider call, so a denied action never reaches a model.
- Acceptance tests decide whether a run shipped something correct rather than something plausible.
- Evals run as a golden suite in CI and block the release when they regress.
- One trace ID connects the request, every tool call, and the receipt.
- Budget ceilings are enforced before the call, not reconciled after the invoice.
- Approval gates suspend the run until a named human approves the proposed action.
- Every write has a documented reverse, so a bad run is reversible.
The patterns
Acceptance tests for AI agents: define done before the run →
Define done before the agent starts.
LLM evals in CI: a golden suite that gates releases →
A golden suite in CI, not a spreadsheet.
MCP server permissions: deny by default, declare every tool →
Deny by default, declare every tool.
AI agent observability: one trace ID from request to receipt →
One trace ID from request to receipt.
LLM cost controls: budget ceilings enforced before the call →
Ceilings enforced before the call, not after the invoice.
Human approval gates for AI agents →
Nothing downstream executes before a human approves.
Agent trace replay and diffing →
Replay the run, diff the divergence.
Durable checkpoints and recovery for agent runs →
A crash resumes; it does not restart.
Rollback for agent-made changes →
Every write has a documented reverse.
Handing over an AI system your team can operate →
The team operates it after the engagement ends.
Where it is implemented
ClawKeeper →
Governed Claude agent control plane. A deterministic policy engine runs before every LLM call, autonomous spend defaults to zero, and the audit log is enforced by the database.
TechTide Harness Kit →
Trust contracts for agent skills expressed as JSON Schema: risk tiers, tool permissions, deny-by-default egress, approval gates, and prompt-injection fixtures behind a wall of CI validation gates.
CipherClaw →
Deterministic multi-agent debugger with zero LLM calls in the analysis path: causal graphs, weighted root-cause scoring, and failure prediction.
Swarm 357 →
357 Claude agent roles across six business layers, with portable memory, enforced cost caps, a bash policy gate, and durable checkpoints.
FintheFinder →
Deep-research agent whose runs suspend until a human approves, with a deterministic citation audit that keeps AI out of the verification path.
TechTideAI2 →
Company-scale agent platform on a three-plane control/data/evidence architecture, with state-machine transitions, hard execution timeouts, and an adversarial eval harness over a 33-task golden suite.
This page documents how agent governance is built. Engagement scope and pricing live on TechTide AI.
AI Production Systems at TechTide AI→© 2026 Alex Cinovoj · TechTide AI · Columbus, OH