From an error report to a reviewed fix
One worked story: an error tracker calls a workflow's webhook, a triage run checks the error ledger and files an item, a fix loop runs until checks pass, and a person approves.
Blog
Lessons from building software with coding agents: what they need to know, how they should work, and where a person decides.
Who decides: your questions and the agent's
Agents stall when they ask you everything and drift when they decide everything; split questions into business and technical ones, and give each kind a fixed path.
When your agent reads 100k tokens of rules before it starts
Rules loaded up front for every task can fill a large share of an agent's context before work begins; load them per task, keep them short and move the rest into checks.
Rules agents follow: checks over prose
A rule written in prose is advice an agent may skip or forget; a check that fails with a clear message is a rule it follows, so move every constraint you can into one.
Letting agents work overnight: the morning review queue
Agents can work while you sleep if the night's work is ready, runs wait instead of guessing, checks decide, and the morning brings one short queue of decisions.
Several agents at once: isolation is the easy half
Separate working copies let several coding agents run at once; the hard half is choosing what may run together, merging it and keeping one person able to review it.
What an agent remembers between sessions
A coding agent starts each session knowing nothing of the last one, so what carries over is only what was written down and handed to it: decisions, not transcripts.
From prompts to loops to graphs
Working with coding agents moved from single prompts to agent loops to workflows shaped as graphs, and each shape answers a limit of the one before it.
Specs for agents: discipline, not waterfall
Writing specs before coding agents build can feel like a return to waterfall, but short specs per item, written just before the work and carrying their own definition of done, keep fast agents on the right problem.
What one finished feature costs in tokens
Agent bills arrive per session or per month, but the number worth knowing is what one finished feature costs, rework and the runs that never shipped included.
Reviewing code nobody wrote
When coding agents write most of the code, review becomes the bottleneck and there is no author to ask; review the plan first, let checks go before people, and judge the result against a written goal.
Why coding agents invent their own words
Coding agents pick the word that sounds most natural rather than the one your product uses, and the drift spreads through code, docs and search until a glossary they check stops it.
One set of rules for every coding agent
Each coding agent reads its own instructions file, so projects that mix agents keep several drifting copies of their rules; keep one set instead and hand each task the part that applies.
The context layer and the workflows for agent-built software
What a context layer and workflows are, why software built by coding agents needs both, and how the two fit together around your code.
Agents write the code. Who keeps the why?
Coding agents now write most of the code, but what a product must do and why rarely lives where an agent can read it, and that gap costs more than slow typing ever did.