Six things we believe about agents that we used to be wrong about.

Two years of shipping multi-agent systems has changed our minds about basically everything. Here are six beliefs we now hold tightly that we used to dismiss — and one we still argue about internally.

1. Bigger context windows aren’t a substitute for memory.

We tried solving every memory problem with longer contexts. It worked until it didn’t — long contexts hide where information lives, and they cost more to debug than to redesign. Memory is structured. Cram less into the prompt; persist more.

2. RAG isn’t a feature, it’s a tool call.

Retrieval is a tool the agent decides to call when it needs something. We stopped pre-stuffing prompts with ‘relevant context’ and started letting agents call retrieval on demand. Better answers, smaller prompts, fewer surprises.

3. ‘Autonomous’ is a marketing word.

Customers don’t want autonomous. They want reversible. They want to know what the agent did, why it did it, and how to undo it. ‘Autonomy’ on its own is a liability surface. Reversibility is a feature.

4. Specialised agents beat one big agent.

Every time we’ve been tempted to ‘just make the model smarter,’ we’ve regretted it. A team of small specialists is faster, cheaper, and easier to audit than one model trying to be everything.

5. Demos lie. Production tells the truth.

Demo runs are short, well-prompted, and curated. Production runs are long, messy, and adversarial. Build for production traces, not for the demo deck.

6. The orchestrator should be boring.

Cleverness in the orchestrator is debt. The orchestrator’s only job is to pick the next agent and route the typed payload. Save the cleverness for the specialists.

“Demos sell the product. Production traces sell the trust.”