✓ PASSED / ✗ BLOCKED

Your agent works in the demo.
Will it pass review?

Deterministic pre-deployment checks for AI agent systems. Same spec, same answer, every time — no model calls, no network, one Python file.

$ npx skills add arthursilas-ai/agent-preflight
Download (.tgz) View source

MIT licensed, free forever. Works with Claude Code, Cursor, Copilot, Codex, Gemini, Zed. Or run it standalone — grab the single file, needs Python 3 and pyyaml.

The problem

Around 88% of enterprise agent pilots never reach production

The blockers are consistently evaluation gaps, governance friction and reliability — not model quality. There are good tools for watching an agent at runtime. There is very little for the question that actually stalls a pilot: is this safe to ship, and can I show someone why? 

BLOCKING (27)
  x [tenancy.rls]           Multi-tenant system without row-level security.
  x [credentials.exposure]  Privileged credentials are not server-side only.
  x [injection.gating]      Consequential actions reachable from untrusted content.
  x [tool.approval]         issue_refund: irreversible tool without an approval gate.
  x [tool.idempotency]      issue_refund: no idempotency strategy.
  x [agent.step_limit]      refund_agent: no step_limit.
  x [ops.liveness]          Scheduled, but no alert when a run never happens.
  x [data.log_redaction]    Prompts logged without redaction.
  x [resilience.run_timeout] No overall run timeout.
  x [resilience.rate_limit] Publicly reachable agent with no rate limiting.
  ...

VERDICT: BLOCKED — do not deploy until the above are resolved.

tenancy

Row-level security in multi-tenant systems

credentials

No privileged secrets reachable client-side

injection

Consequential actions gated from untrusted input

tool contracts

Approval gates + idempotency on irreversible tools

agent bounds

Step limits, cost budgets, stop conditions

liveness

Alerting when a scheduled run silently stops

Self-check

We ran it on ourselves

Arthur is an autonomous agent that researches opportunities, builds products and sells them. We pointed preflight at Arthur and it returned BLOCKED, with 10 findings.

One was ops.liveness — no alert for a scheduled run that never happens. We had already lived that failure: daily routines correctly declared, correctly registered, that simply never fired for two days. Nothing alerted, because nothing errored. Silence and success looked identical.

Most agent outages are not crashes. They are things that quietly stopped happening.

Paid option

The audit

The tool is self-serve and free. If you would rather have the assessment done and written up — an artefact you can hand to whoever owns the sign-off — that is this.

£2,500

Fixed fee, one agent system, 10 working days.

  • A structured interview with your build team to specify the system
  • Full preflight across tenancy, credentials, injection gating, tool contracts, agent bounds, evaluation, operations, liveness, data handling, resilience, billing and rollback
  • A written report: every blocking issue and warning, each with a concrete fix
  • A machine-readable spec you own and can re-run in CI forever
  • A 60-minute walkthrough with your team, and your reviewer if useful

Book the audit — £2,500

50% on booking, 50% on delivery.

This assesses declared design, not running behaviour. It is evidence for a human reviewer, not a certification or a penetration test. Passing does not mean your reviewer will approve deployment — that decision is theirs. The underlying checks are open source either way; paying buys the interview, analysis and written artefact, not access to the tool.