Architecture review

A structured design review for a system where software takes actions on its own — a multi-agent framework, an orchestrator-plus-workers setup, an LLM-in-the-loop automation, a runbook-automation or SOAR build. You provide the design; the deliverable is a report that tells you where the trust boundaries are thin, what could go wrong and how badly, and what to change before you widen autonomy.

It is the same analysis that produced the architecture pages on this site, pointed at your system instead of a hypothetical one.

• The reversibility line — which actions are safe to automate (reversible, bounded scope) and which must stay with a human, and whether your design draws that line where the blast radius actually is.
• The human gate — is there one, is it staffed, is an approval request self-contained, does it have a TTL, and can it be bypassed under load or during an incident?
• Least privilege & credentials — standing vs just-in-time access, credential exposure on the wire or in logs, rotation, and what a single compromised agent can reach.
• Audit & state — is there an append-only record ahead of every action, is it external to what the agents manage, and is system state a projection of it or a second source that can disagree?
• Failure handling — timeouts vs hangs, idempotent retries, circuit breakers, dead-letter handling, poison messages, and partial failure as a first-class outcome.
• Blast radius & the deny-list — the worst realistic bad run, and whether there is a fixed set of actions no tier and no prior approval can authorise automatically.
• Rollout & operations — is autonomy widened on evidence or on optimism, are there kill switches, are they tested, and are prompt and model changes run through change management?

A written report, typically 8–15 pages, structured as:

1. Summary & readiness call — a one-page verdict: ready to widen autonomy, ready with fixes, or not yet, and why.
2. System as understood — the design restated in the same shape as the architecture pages here, so any misread is caught before the findings.
3. Findings, ranked by risk — each with the scenario that triggers it, the realistic impact, and a specific recommendation. Not a checklist score.
4. The trust-boundary map — where a human is (and isn't) in the loop, drawn against where the irreversible actions are.
5. A rollout adjustment — a concrete phased path from where you are to broad autonomy, with exit criteria per phase.
6. Open questions — what could not be assessed from the material provided.

One round of written follow-up questions on the report is included.

1. You email a short description of the system and what you want out of the review.
2. Scope & price agreed in writing — a fixed fee for a defined deliverable and turnaround, no hourly meter.
3. You send the material — design docs, diagrams, prompt/ policy files, a written walkthrough. Redact anything sensitive; the review does not need secrets or live access.
4. The report comes back within the agreed window (usually about a week for a typical system).
5. One follow-up round of written Q&A on the findings.

• Not a security audit, penetration test, or certification — it is a design review, and it says so in the report.
• Not an assessment of your running systems — no credentials, no network access, no agent connected to your estate. Paper only.
• Not an automated or instant product — every review is arranged, scoped, and delivered by arrangement. If you want the self-serve version, the full-edition guides and the free architecture pages cover the same principles.
• Not a compliance sign-off. It will tell you where your design diverges from the reversible-first, human-gated model these pages describe; it does not map to a specific framework's control catalogue unless that is agreed in scope.

Email apacheshadow1972@gmail.com with a few lines about the system and what you want assessed. You will get back a scope, a fixed price, and a turnaround before anything is committed.

Prefer to self-serve first? The agent operations playbook, the service-desk and SOC architectures, and the coordination protocol are all free and cover the checklist above in full.