How do agents go from demo to production?

FFDE draws on 12-Factor Agents: agents stuck at demo quality usually fail on engineering — prompts, context, control flow, and state hidden in framework black boxes. Production-ready delivery means mostly deterministic software with LLM steps at the right points; own prompts and context, keep agents small, pause for human approval on high-stakes steps, and make runs replayable — aligned with Spec/Eval/Handoff and the three decision gates.

Key points

  • Core question: why can your agent only run demos, not production? — the 70–80% quality wall needs engineering, not another framework
  • Three foundations: LLM as pure function (context engineering) · Agent as software · Own your control flow
  • One Agent MVP per use case; main path ~3–20 steps — context, tools, state, errors, control flow
  • Version prompts and RAG/context strategy in an Agent Package — diffable and rollback-ready
  • Research/query steps may loop; writes, external publish, and deploy must break for human review
  • Human approval is a first-class tool — maps to intent, scope, and GO/NO-GO gates
  • Unify workflow/orchestration runs with business events for audit and evidence packages
  • Golden eval sets include failure cases; drop resolved errors from runtime context
  • Trigger from WeChat, email, cron, webhooks — meet users where they work

How to do it

  1. Stage 1: freeze Top 1 use case and eval criteria (Focused FDE)
  2. Stage 3: classify steps as sync continue / pause / must-have human review
  3. High-stakes actions default to preview + manual review — no auto-commit
  4. Stage 4: golden eval + five-dimension ROI + signed GO/NO-GO

Related reading

Book a free consultation →