Your demo worked. Your agent is still not production-ready.
A demo is one good run on one chosen input. Production is thousands of runs on inputs nobody chose, with side effects you can't undo by refreshing the page. Yet teams keep shipping agents the moment the demo stops embarrassing them — and then act surprised when the thing that aced the boardroom demo starts refunding customers, deleting records, or looping itself into a $400 bill at 3am.
Readiness is not a launch date or a feeling of confidence. It's a checklist you can measure, defend, and roll back from. Here's the take.
Stop grading your agent on the leaderboard
The primary readiness signal is a task-success rate measured on your work, not a benchmark leaderboard. Build your eval suite from real failures, written the way real users write them — Anthropic's own guidance says start with just 20 to 50 tasks from real failures.
And measure reliability across repeated trials, not single runs. A 75% per-trial success rate compounds to roughly 42% over three runs. The tau-bench authors found state-of-the-art function-calling agents succeeding on fewer than half of retail tasks, with pass^8 below 25%. Your agent's one good run proves nothing. Its tenth run tells you the truth.
Guardrails belong in code, not in the prompt
Telling a model "never delete records" is guidance. Configuring its permissions so it cannot call any record-deletion API is a guardrail — and it holds even when the model gets malicious input or just has a bad day.
Every action gets a risk tier. Read-only actions run free. Refunds, payments, deletions and contract changes always need approval or strict rule-based thresholds. And human-in-the-loop doesn't mean a human reviews everything — it means a human is in the loop for the decisions that warrant it, with earned autonomy as the agent proves itself case type by case type.
Hard limits are part of the same layer: cap steps, tokens, spend and wall-clock time per task, and give the team a kill switch anyone can trigger without a deployment. When a limit is hit, the agent stops and escalates — it doesn't try harder.
If you can't see it failing, you're not ready
Weak observability and immature guardrails were the most common pain points in Cleanlab's 2025 production survey — and over 65% of teams listed improving observability as their top investment priority. Full traces, dashboards and alerts on cost, latency and failure-rate drift are table stakes, not nice-to-haves.
Run continuous evaluation on sampled production traffic, and when a failure appears, turn its trace into a regression task so the same failure gets caught automatically next time. Read the transcripts before trusting the scores: a 0% pass rate usually means a broken task or grader, not a broken agent.
Ship in gates, or don't ship at all
Shadow mode first — mirror live traffic, discard the outputs, zero blast radius. Then a canary at 1–5%, then 5%, 25%, 50%, each behind explicit numeric gates written down as configuration, not agreed in a meeting. Rollback must be mechanical: triggers in config, previous version kept warm, and — the detail everyone forgets — semantic caches invalidated, or they'll keep serving bad answers for hours after you've rolled back.
The most common mistake is shipping on demo prompts instead of production traffic evals. The second is launching with no rollback plan and nobody watching. The third is treating a single good eval run as proof. None of this means agents should stay in pilot forever — Klarna's service agent now handles two-thirds of chats and cut resolution from 11 minutes to 2. That's what a production agent looks like after it has passed its gates. The question isn't whether agents can be ready. It's whether yours has earned it.
Go deeper
Read the full readiness checklist — eval thresholds, guardrail design, the staged rollout plan and an example release-gate configuration — on Agentic Affairs:
👉 When Should AI Agents Be Considered Ready for Production Deployment?
Originally published on Agentic Affairs (https://agentic-affairs.com)
Top comments (0)