Agent Lifecycles Should Be Model-Shaped, Not Human-Shaped

·https://www.voodootikigod.com/adlc-1-models-arent-human
articleaiagentic-developmentagent-loopssdlctestingcode-review

maps directly to joelclaw agent-loop role separation: test writer, implementor, reviewer, and judge gates should defend against model failure modes instead of copying human SDLC theater

Chris Williams has the cleanest framing here: the software development lifecycle is mostly a defense system against human bullshit. Standups, reviews, estimates, docs, handoffs — all useful, but all shaped around human failure modes like ego, fatigue, forgetting, politics, and fear.

That gets weird when teams copy the org chart into agents. A product manager agent, a senior engineer agent, a reviewer agent, a tiny fake standup, and then everyone acts surprised when the system ships a convincing storefront with stubbed data behind the counter. The ritual looked familiar, so it felt safe. It wasn’t.

The clever move in Agentic Development Lifecycle is the rule: every phase, gate, and loop has to trace to a model failure mode it defends against, or a model property it exploits. That turns process design into something closer to threat modeling. Premature satisfaction, sycophancy, context rot, confident hallucination, Goodhart’s Law, review-count priors, generative bloat, and cross-session coherence loss each demand different rails than human teams do.

This rhymes hard with joelclaw agent loops: separate contexts, executable acceptance criteria, deterministic evidence, protected tests, fresh reviewer passes, and a judge that doesn’t care how charming the builder sounds. Less fake Agile cosplay. More machine-shaped lifecycle.

Key Ideas

  • Chris Williams argues that the SDLC was built around human failure modes, so directly copying it into agentic development imports the wrong defenses.
  • The core design rule is that each agent lifecycle phase should defend against a specific model failure mode or exploit a specific model property.
  • Model failures named in the article include premature satisfaction, sycophancy, context rot, confident hallucination, Goodharting, review-count priors, generative bloat, and cross-model coherence loss.
  • Useful model properties include cheap sampling diversity, aimable sycophancy, no ego, no fatigue, fresh-context review, and a cost model where exploration and rewriting are cheaper than human coordination.
  • The lifecycle implications map cleanly to joelclaw patterns: fresh creator/critic contexts, executable tests as rails, deterministic gates, protected acceptance criteria, fan-out review, and post-merge simplification.

Links

← All discoveries