Production Agent Patterns: What I Now Default To

Every new production agent I build now starts from a pattern library. Not from a blank prompt. The library was earned, one bug at a time.

Default 1: Bounded Scope per Run

No agent gets an unbounded task. Every run has a declared input, a declared output, and a time budget. Unbounded agents either stall or run wild. Bounded agents produce predictable cost curves.

Default 2: Idempotent Operations

Every write operation the agent can execute is idempotent by design. The same input produces the same output. Re-runs are safe. This is non-negotiable because production agents retry more than production humans.

Default 3: Explicit State Machine

The agent runs inside a state machine: pending, running, blocked, completed, failed. Transitions are logged. The coordinator or human can see where any agent is without reading logs. This replaces vibes-based debugging with state-based debugging.

Default 4: Structured Outputs Always

The agent returns JSON with a schema. Not prose. Even for human-facing results, the prose is a field inside structured output. This lets downstream code consume, log, and retry without string parsing.

{
  'status': 'completed',
  'summary': 'Renamed 14 files and updated imports',
  'artifacts': ['src/...'],
  'next_action': None,
  'errors': []
}

Default 5: Cost Budget Per Call

Every agent call has a max token budget. If the budget is exceeded, the agent stops and reports. No silent token blowups. Over the year this alone saved thousands of dollars.

Default 6: Human-Readable Trace

Every production agent writes a human-readable trace file that a non-technical stakeholder can read. Not a debug log. A narrative of what happened. This is how I explain agent failures to clients without hand-waving.

Default 7: Graceful Abort

Every agent listens for an abort signal and cleans up mid-run. Locked files released, partial commits reverted, in-flight calls cancelled. If your agent cannot gracefully abort, it is not production-ready.

What I Do Not Default To

Autonomous multi-step reasoning. Self-modifying prompts. Cross-agent negotiation. Tools that can delete without a dry run. These can work, but they are not defaults. They are advanced options for specific problems.

The Anthropic production guide spells many of these out. Read it twice.