The orchestrator is not Claude
On the very second day I hit a fundamental architectural decision: the orchestrator must be separate from the model that does the coding. Claude (or any other language model) is a tool of the orchestrator — not the orchestrator itself.
The orchestrator is the logical layer that holds the project state, plans steps, chooses the right model for each task and verifies results. In practice this means the orchestrator can switch from Claude to another model without changing anything in the workflow.
This decision proved crucial on day four — but I'm getting ahead of myself.
Phase 0: before the agent starts coding
Every work cycle starts with phase 0. The agent doesn't go straight to code — it first reads the context: relevant files, existing tests, architectural decisions. Only then does it propose a plan.
This phase is expensive in tokens, but cheap in mistakes. An agent that skips phase 0 and jumps straight to code creates things that don't respect existing conventions. Cleaning that up costs more than phase 0 itself.
15 minutes AI, 45 minutes human
The actual ratio on day two was surprising: the agent was actively working for roughly 15 minutes out of every hour. The rest of the time I spent reviewing output, making architectural decisions, writing context for the next step.
This isn't a failure of the agent. It's the correct division of labour. The agent is fast at implementing concrete steps — patterns, boilerplate, data transformations. It's slow on decisions that require business context or a view of the whole architecture. That's where the human is irreplaceable.
When measuring productivity, don't count just the agent's time. Count the total time per feature — and that was significantly lower than without AI.
Git as a checkpoint
On day two I introduced a rule: every completed agent step = a commit. Not at the end of the day, not after a feature is done — after every logical block of work.
The reason is simple: the agent can go wrong. With granular commits, rollback is trivial. Without them you're manually digging through diffs trying to figure out what the agent actually changed.
The complete project repository is available at: https://lnkd.in/dutnJmBQ
What validation delivered
At the end of each work cycle the agent ran a set of tests. Not the full test suite — that would take too long. Targeted tests covering the parts of the code it had just modified.
Result: three out of five cycles passed on the first try. Two required a fix. In both cases the agent itself identified the problem and corrected it without my intervention. This is the moment the agentic loop starts making sense.
Want to see how business processes can be automated? Book a consultation — we start where vibe-coding ends.
In the next episode
Day 3 brings numbers that might surprise you: what agentic AI development actually costs and how it compares to a senior developer.