Five Production-Ready Elements
Boundaries, Signal Tuning, Error Handling, Tool Handling, Model Routing. What separates toys from production.
You can build an agent that works in a demo. Making one that works in production — reliably, day after day, without your supervision — requires five specific elements. Miss any one of them and you'll spend more time babysitting your agents than they save you.
1. Boundaries. You've been writing these since Stage 1, but production boundaries are more rigorous. They don't just say what the agent should do — they define the exact scope of operation with zero ambiguity. "Process invoices from the /incoming directory. Move processed invoices to /completed. Flag invoices over $10,000 for manual review. Do not process files from any other directory. Do not delete any files."
2. Signal Tuning. Your agent receives a lot of information. Signal tuning is about helping it focus on what matters and ignore what doesn't. In your system prompt: "Focus on: error messages, status codes, file changes. Ignore: deprecation warnings, informational logs, comments in code." Without signal tuning, your agent processes noise and wastes tokens on irrelevant details.
3. Error Handling. Production agents encounter errors. Network failures, missing files, malformed data, API rate limits. Every error scenario needs a defined response. The golden rule: never fail silently. If an agent can't complete a task, it must report what happened, what it tried, and what needs manual attention. Silent failures are production killers.
4. Tool Handling. In production, tool usage needs guardrails beyond what's in your development setup. Timeout limits for bash commands. Retry logic for web requests. File locking for concurrent access. Validation of tool outputs before using them in the next step. These small details prevent the cascading failures that take down production systems.
5. Model Routing. Not every task needs your most expensive, most capable model. A scanning task that reads 50 files and extracts dates? That's Haiku-tier work. A synthesis task that compiles findings into a strategic recommendation? That's Opus-tier. Production systems use the right model for each task — saving 50-80% on API costs while maintaining quality where it matters.
Walkthrough
- 1Take your most-used workflow and audit it against all five elements
- 2Boundaries: can you describe the exact scope of every subagent in one precise paragraph?
- 3Signal Tuning: add 'Focus on' and 'Ignore' sections to each subagent prompt
- 4Error Handling: for each subagent, list the 3 most likely failure modes and define the response for each
- 5Tool Handling: add timeout and retry instructions for network-dependent tools
- 6Model Routing: identify which subagents need high-capability models and which could use cheaper ones
- 7Document your findings: which elements were weakest in your current setup?
- 8Create a production readiness checklist based on the five elements
Checkpoint
You've audited your workflow against all five production elements and identified specific improvements for each. You have a written production readiness checklist.