Agentic SDLC Patterns

Public Posts

• The Day Two Agents Drafted the Same Six Posts

  Two agent sessions independently drafted the same wave of six blog posts at once. Nobody planned that. The fix was not to prevent it, it was to make it survivable.

• A BLOCKED Banner as an API Between Two Labs

  Three real runs failed on the same trivial bug, and every new session wanted to try again. The fix wasn't a retry. It was a sentence in the backlog: do not pull on this thread again from this side, and here is exactly what reopens it.

• Two Sessions Claimed RUN-0023

  Two agent sessions closed out in parallel, both read the same last id, both wrote RUN-0023. Then it happened again with RUN-0026. Sequential ids in a multi-writer log are a collision generator by design, and the ledger sitting next door proves the alternative.

• One Blind Spot, Found Twice

  In two weeks across four labs, a second AI reviewer caught a fabricated example, rescoped a whole system, confirmed the same bug twice independently, and flagged a security check testing the wrong path. One thesis, five exhibits: independence is what you're paying for.

• Review the Intent, Not the Diff

  Agents produce code faster than one human can read diffs. Intent-assurance-lab's bet is to move the review unit up a level: a falsifiable statement of what a change should make true, plus the evidence, with line-by-line reading reserved for where that evidence is thin.

• Don't Write Intent for Code That Already Exists

  Intent-assurance-lab's first real trial needed a genuinely unstarted task. The obvious candidate turned out to have live, uncommitted work already in progress on the exact feature -- and writing intent against it would have proven nothing except that hindsight writes clean specs.

• Status Is Derived, Never Stored

  The original design had a status field you edited. The rewrite has no status field at all, just an append-only event log and a pure function that replays it. Concurrency safety and text-editor repair both fall out for free.

  Part 2 of 9 in From Control Tower to First Product
• Build the Fake Project First

  Every remaining feature got proven against a checked-in, disposable demo project before touching anything real, until an ARB review found the demo's own success report could go green without ever actually checking the thing it claimed to prove.

  Part 6 of 9 in From Control Tower to First Product
• Verifying the Wrong Universe

  Worktree isolation was correct. Post-task verification was correct. Composed together with no validation at the seam, a relative path silently pointed the verify step at the wrong directory, and produced a confident, meaningless result.

  Part 8 of 9 in From Control Tower to First Product
• Teach the Engine to Decline

  Pass/fail is not enough vocabulary for an autonomous engine. A third outcome, declined, lets a caller escalate automatically through a whole chain of engines and still refuse cloud egress the instant no human is there to consent to it.

• One Human Sentence, Two Agents, Two Repos

  A converged run still had to be closed out across two repositories. The human's entire contribution was one approval sentence in chat. Everything after it was two agents in two repos completing a live handshake, neither one ever committing in the other's tree.

• No Model Reviews Its Own Homework

  Four documented review cycles, one recurring shape: a fix, left unreviewed by a second perspective, quietly introduces the next bug. The sentence the factory arrived at after proving it the hard way, four times.

  Part 1 of 6 in No Model Reviews Its Own Homework
• The Fixes That Held Changed a Rule

  Three adversarial review rounds on the same new codebase, and the diagnostic that separated a durable fix from a fragile one: did it change a rule, or just a line?

  Part 2 of 6 in No Model Reviews Its Own Homework
• Doubt Before the Code Exists

  Adversarial review, applied at intent-declaration time instead of code-review time, caught real defects before a single commit existed to review, proving the doctrine is a decision-making technique, not just a code technique.

  Part 3 of 6 in No Model Reviews Its Own Homework
• It Wasn't a Matched-Cast Comparison

  A live-model coordination comparison looked decisive: peer-delegation succeeding, central-orchestrator starving the colony. An independent review asked whether it was actually a fair fight. It wasn't.

  Part 4 of 6 in No Model Reviews Its Own Homework
• Shipping With a Watch, Not a Verdict

  Five doubt cycles on one small registry guard, and the finding rate never reached zero. What actually caught the headline bugs was mutation testing and fresh-context review, self-inspection caught none of them.

  Part 5 of 6 in No Model Reviews Its Own Homework
• Two Reviewers, One Exploit, Two Real Bugs

  Two AI reviewers, independently, with real build and openssl access, found a bypassable revocation path and a spoofable certificate identity field in device-CA code. Both were confirmed with a live exploit, before and after the fix.

  Part 6 of 6 in No Model Reviews Its Own Homework

Draft Queue

No draft placeholders assigned to this series yet.