Reliable autonomous coding agents come from architecture — deterministic stop signals, isolated context, programmatic gates — not smarter models.
Stop reason is the only termination signalParsing text for "I am finished" fires on "I am finished with step one" and crashes the migration halfway through.
Code enforces rules, prompts only suggest themAn all-caps prompt ban on drop table works 99% of the time; a prerequisite gate that intercepts the SQL works every time.
Narrow scope beats big context for subagentsEighteen tools instead of four or five explodes decision complexity, and 40 turns of architecture debate makes a unit-test agent try to redesign the database.
Executive SummaryAI
The hosts open on two outcomes of handing an AI your production database — a flawless pull request or a recursively deleted billing module — and argue the difference is not the model's intelligence but the architecture around it, then set up a fragile legacy Python monolith called Paystream as the running migration example for this Claude Architect exam study guide.
The agentic loop must be driven entirely by the stop reason field: a tool_use pause means your code runs the tool and appends the result to the conversation history as a new message, end turn ends the loop, and parsing assistant text for phrases like "I am finished" or using an iteration cap as the primary stop (a 50-iteration failsafe is only a backstop against API bankruptcy) are the fastest way
For a monolith this size one agent drowns in context, so the coordinator–subagent hub-and-spoke pattern spawns narrow workers through the task tool — each waking up blank with only its system prompt and explicitly passed structured context, up to five in parallel in one response, with no peer-to-peer whispering and iterative re-delegation when the coordinator finds gaps.
Hard rules like never running drop table cannot live in the prompt — prompt instructions are probabilistic — so you build programmatic prerequisite gates with agent SDK tool-call interception hooks to block violations before they reach the database, and post-tool-use hooks to normalize messy output (Unix timestamps, ISO 8601, a proprietary 90s numeric string) with cheap Python before the model bur
Over a six-hour run the fix for the lost-in-the-middle effect is not progressive summarization, which smooths away exact table names, column types, port values and error codes, but extracting transactional facts into a persistent case facts block injected at the top of every request, plus subagent scratchpad files, JSON state manifests for crash resumption, and stratified sampling of high-confiden
Key Quote
“You can, and honestly, it'll work 99% of the time.”
— Expert
Key Quote
“Yes, that description is the only user interface the model has.”