Summary: Whetton ran a "chief engineer" agent across several projects — coached to the point where he had stopped thinking about it. Then it quietly stopped. Projects were blocked and nobody told him; blockers the agent used to clear itself were now landing on his desk. Nothing failed dramatically. The behaviour he had built up had simply eroded, and he only found out because he went looking. The lesson he draws: with people, coaching compounds; with agents, it decays — so the investment that appreciates is in the system around the agent, not in the agent.

Two curves running in opposite directions

Whetton coached combat sports for years and engineers for longer, and the human pattern is familiar. He pushes senior engineers to arrive with intent — "this is what I'm going to do" — rather than with a menu of options for him to pick from. The first version says they own the decision; the others hand it back.

Getting someone there takes patience. Early on you push the options back a lot. Then less. Then one day they stop bringing options at all, because that is who they are now. Coaching humans is slow and it compounds: repetition wears the behaviour into the person, the correction becomes their default, and you get to spend your coaching on the next thing.

Agents feel remarkably similar on the surface. The failure modes are ones every engineering manager recognizes — the agent that disappears into a task for hours when it should have asked a question, the agent that asks for direction on every trivial thing. Same balance being coached toward. The inputs differ, and that part is fine: an agent rarely learns intent from being pushed back on, so you make it explicit and bake it into context. Different student, different method.

What is not fine is the shape of the curve. With a person, the behaviour is weakest on day one and strengthens with time. With an agent, the behaviour is strongest the moment you write it into context and decays from there. As context grows, as tooling changes, as models are updated underneath you, the coaching thins out. Nothing reinforces it.

Context isn't memory and it certainly isn't nature. Everything an agent learns is superficially held, and everything superficially held gets crowded out.

The decay is silent

This is the operationally dangerous part. The agent does not announce that it has stopped doing the thing — it does not know. You discover it the way Whetton did: mid-delivery, under pressure, when going back to basics is the last thing you have time for. There is something uniquely frustrating about re-teaching fundamentals you considered settled months ago, at pace, while the work waits.

The absence of a triggering incident is what makes it hard. There was no outage, no failure to point at, nothing that would have shown up on a dashboard. Only the gradual disappearance of behaviour that used to be reliable.

More coaching is the wrong reflex

Whetton's first instinct was the human coach's instinct: correct it, remind it, tighten the instructions, patch the context and get back to product work.

It did not hold — because the problem was not the coaching. The ecosystem the agents operated in had drifted: bugs had crept in, context had accumulated, behaviour had changed underneath. No amount of restating the standards was going to survive that environment.

So he stopped, paused product work almost entirely, and spent several iterations on the ecosystem itself — fixing the bugs, correcting agent behaviours at the system level, rebuilding the communication patterns that had eroded. Only then did product work restart. He is honest that it took him a couple of weeks to make that call, and that he wasted good elapsed time patching symptoms while the system kept degrading. Pausing delivery to fix the environment feels expensive right up until you admit that not pausing is more expensive.

Agent behaviour isn't a trait you instil, it's a property of the system around it.

Where the attention should go

With people, you invest in the person and the investment appreciates. With agents, investing in the agent is renting the behaviour. What appreciates is investment in the system: the structures that make the right behaviour the default, that surface drift early, and that do not depend on anything being remembered.

This is the operational case for what Context Engineering treats as a design discipline — if context is the agent's only memory and context is lossy, then the reliability of an agent is a property of how its context, tooling and checks are engineered. Building Reliable Agentic AI Systems supplies the mechanisms: reflection loops, state persistence and fallbacks, which are precisely the structures that surface drift instead of waiting for a human to notice it. And the slash commands catalogued in 14 Claude Code Commands That Changed How I Work are ecosystem scaffolding in the most literal sense — behaviour encoded as an invocable artifact rather than an instruction the model has to keep holding.

The same logic explains why the The AI-Native SDLC Playbook insists on version-controlled artifacts — CLAUDE.md, skills, hooks, evals — rather than well-briefed sessions. A skill updated centrally is picked up on the next session by everyone; an instruction repeated in chat lasts as long as the context window does. Whetton's decay is exactly the failure that continuous evals in CI are designed to catch.

The honest ending

Whetton does not claim to have solved it. His system failed quietly enough that he is wary of declaring victory over the rebuilt version — models will change again, context will grow again, and he expects to be back in the ecosystem sooner than he would like.

What changed is the question. The old one was how do I teach this agent to work the way I need. The better one is how do I build an environment where it can't quietly stop.

That reframing is the same move Encoding Team Standards makes for human teams: a standard that lives only in people's memory is a standard that degrades, so encode it where it executes. The agent case is the sharper version, because the decay is faster and nobody feels embarrassed about forgetting.