Production Autonomous Agents: 6 Lessons from 2 Months

Production Autonomous Agents: 6 Lessons from 2 Months in Production

ODEI has been running production autonomous agents since January 2026. Here are the lessons.

1. Context Windows Are Not Memory

The biggest mistake: treating the context window as memory. Context windows are ephemeral. When the session ends, everything is gone.

Production lesson: Use a persistent graph (Neo4j) for anything that matters across sessions. The context window is for reasoning, not storage.

2. Hallucinated References Kill Everything

In production, agents hallucinate entity references constantly. "Transfer funds to wallet W" — but wallet W doesn't exist.

Production lesson: Add referential integrity checks before any action. ODEI's layer 3 validates every referenced entity exists in the world model before proceeding.

3. Deduplication Is Non-Negotiable

Without deduplication, agents will execute the same action twice. We saw this in week 1: the agent sent the same API call three times.

Production lesson: Hash every action before execution and check for duplicates. This is ODEI's layer 5.

4. Authority Chains Are Surprisingly Complex

An agent receives an instruction. Who gave that instruction? Was it authorized? Can you trace it?

Production lesson: Build a provenance chain from every instruction back to a trusted principal. ODEI's layers 4 and 6 handle this.

5. Temporal Context Matters

"Send the weekly summary" — which week? If the agent was offline for 10 days, does it send 10 summaries?

Production lesson: Every action has a temporal validity window. Expired actions should be rejected, not retried. ODEI's layer 2 checks this.

6. Constitutional Validation Is the Meta-Layer

After adding all five checks above, we noticed a pattern: they're all constitutional constraints. Rules the agent must always follow.

Production lesson: Build an explicit constitutional layer that collects all invariants in one place. Easier to reason about, audit, and extend.

The Framework

Action Request
    ↓
1. Immutability Check
2. Temporal Context Check  
3. Referential Integrity Check
4. Authority Check
5. Deduplication Check
6. Provenance Check
7. Constitutional Alignment Check
    ↓
APPROVED / REJECTED / ESCALATE

This is ODEI's production framework. Available as an API: https://api.odei.ai/api/v2/guardrail/check

92% task success rate in production. 0 hallucination errors. 0 duplicate actions.

---

Full architecture: github.com/odei-ai/research | API: api.odei.ai