Teaching Agentic Coding to a Team
Why your playbook won’t work and what to share instead
This week I paired with one of our fastest engineers on a webhook refactor. Quick and dirty job: break a monolith endpoint into handlers, keep the abstraction thin. I pulled up Claude Code and entered planning mode.
We watched the model gather context. Reading files. Building a map of the codebase. Then it laid out the approach before writing anything.
“What’s it doing?”
Planning mode. He’d been using Claude Code for months. Didn’t know it existed.
We kept going. I mentioned trigger words: tokens that nudge the model into specific behaviors. “TDD” gets you tests first. “Tracer bullet” gets you a minimal end-to-end path. “Acceptance criteria” changes how it frames done.
Wide-eyed.
This is someone I assumed was way ahead. He’s fast, ships clean code, stays on top of things. Turns out he’d been using Claude Code like an autocomplete engine.
The Gulf I Hadn’t Noticed
Meanwhile my TL and I are deep into orchestration. I’ve been running Ralph loops; he’s testing a Claude Code binary with swarms. We presented this to the team last week. It went straight over everyone’s heads.
A recent podcast nailed it: across the industry you’ve got people copying from ChatGPT all the way to multi-agent orchestration. Same teams. Same companies. I’d been so focused on squeezing extra percentage points out of my own workflow, I hadn’t noticed my team wasn’t even in the same conversation.
The Instinct That Won’t Work
My first instinct: codify what works. Document the loops. Write up the patterns. Hand the team a playbook.
The reality? The more I codified it, the less other people wanted to use it. My TL doesn’t want to use my Ralph loops and I haven’t downloaded the binary.
Why Resistance Happens
This isn’t laziness. It’s ownership.
A workflow is a set of decisions. Using someone else’s workflow means accepting their decisions without understanding the trade-offs. Why these exit criteria? Why this directory structure? Why JSON over plain text for progress?
I learned those answers by hitting walls. They want to learn the same way.
And honestly? They should.
Someone who hasn’t felt the pain of context overflow won’t understand why you intentionally compact. Someone who hasn’t watched an agent spin for 20 minutes on a solved problem won’t appreciate exit criteria. The playbook assumes experience they don’t have yet.
Right-Sizing to Where People Actually Are
The consistency vs variance trade-off applies to people, not just prompts.
Not yet in the workflow: Start with the simplest integration. “Try planning mode on your next task.” No agents. No automation. Just: does this feel different? Let them discover the value before introducing complexity.
Using AI but unstructured: Introduce one constraint. A CLAUDE.md file. A habit of stating what “done” looks like before starting. Small structure, high autonomy.
Structured but single-threaded: Now you can talk about exit criteria. Backpressure verification. When to let the agent run vs when to intervene.
Ready for orchestration: Full autonomy with shared conventions. They don’t need your workflow. They need your vocabulary.
Most of my team is in the first two buckets. Presenting swarms was like teaching calculus to someone learning arithmetic.
What to Actually Share
Not the specific setup. That’s personal, built from my mistakes. Maybe it always will be personal, for every single developer.
Share the principles. Let them build their own harness.
Factory tolerances: output should be CI-ready. If the agent produces something that needs manual cleanup, your criteria are wrong.
Backpressure verification: start each criterion by asking what the agent can test itself without human input. If it can’t verify it, you’re the bottleneck by design.
When to compact: context isn’t free, but compacting too early loses the reasoning chain. Too late, you end up in the dumb zone of a model. There is no formula. You learn the feel.
They’ll build differently. That’s the point.
What’s your experience teaching AI workflows to a team? Please share if you found this helpful!