Run by agents. Governed by humans. Where human-in-the-loop becomes runtime architecture.
The iteration loop from demo agent to production-ready: validate, read the cited doctrine, fix, re-run. Installable inside Claude Code, Cursor, and Codex via MCP.
Every score links to a public validator run.
Key Facts
- For
- Product designers, PMs, founders, AI builders, and engineers shipping agentic systems
- Handbook
- 10 principle pages connected to implementation examples and runtime architecture patterns
- Learning
- 2 course pages with linked labs and implementation references
- Practice
- 103 implementation examples mapped to principles, runtime patterns, and problem types
Works with
Watch the film
Run by agents. Governed by humans.
Four machine apocalypses, one missing rule. A short film on why governance, not fear, is the answer for agentic AI.
Live readiness review
Two real runs, both inspectable
One we deliberately tried to break, one a real system we validated.
Try it on your code
Public repos · Validator runs on approval · Results in your account
If your agent ships any of these, the doctrine has a name for it
Four recurring failure patterns on agentic systems in production. Each one links to the principle that addresses it.
Case studies
All case studies →Blueprint Readiness Score
58/ 100
Draft
A2A reference agent: what the validator finds in our own example
Single-pass architect.validate run against aidesignblueprint/integrations, the A2A reference example + stdio proxy. 58/D, draft. Four production blockers framed as deliberate scope of a protocol demonstration, one hardening recommendation that rides a small companion fix on the integrations repo. The first case study to publish AUX-pattern annotations explicitly.
Read →Blueprint Readiness Score
22/ 100
Draft
Layer 3: applying the doctrine to claude-agent-sdk-demos
AIDB's cohort-bridge auto-bundled the email-agent SDK glue layer of anthropics/claude-agent-sdk-demos and submitted it to architect.validate. Anthropic publishes these as reference implementations for local development, not production. The validator engaged mechanism-specifically with the substrate the doctrine itself runs on top of. 22/F · high_risk · draft — seven production blockers, P8 (Approvals) at sev 95.
Read →Blueprint Readiness Score
100/ 100
Production-ready
The bridge that selects what feeds the validator — audited by the validator
Fourteen iterations of architect.validate against the cohort-bridge orchestrator that bundles applicant repositories for the validator. 35/F → 100/A. The cert reviewer confirmed the production_ready verdict on iter14: no specific missed defect that would cause silent wrong results, crash, or trust-boundary bypass.
Calculated ROI: $80K – $200K / yr
Read →Latest guides
All guides →
Treat redirection as a state change, not a new prompt.
Restart-heavy UX destroys context and the trust that the agent built up to that point. Steering without restarting is a state change at the run level, not a prompt at the start.
Read →
Make delegated work visible enough to trust, steer, and resume.
Silent agents lose trust the moment the user can't tell what's happening. Async execution must surface its own state: started, progressing, blocked, complete.
Read →
Parallel sessions only scale when ownership, blockers, and merge risk are explicit.
Two sessions, three minutes apart, edited the same row. One silently overwrote the other. Ownership, blockers, and merge risk need to be visible before the second session starts, not after the conflict.
Read →