Don't try the dead end twice

Solo buildersSolo builder who already learned the hard way

We spent an evening trying WebSockets on a serverless host before learning it can't keep connections open. A month later a new session proposed the same approach for notifications.

How do I make my AI agent remember an approach that already failed so it doesn't try it again?

With ArcRoom, the agent records a failed approach with remember_failure, including why it failed. When a later request matches, the failure is handed over with the reason, and the agent picks another route. Only failures someone recorded are known: a silent failure teaches the playbook nothing.

Also asked as: My agent keeps proposing an approach I already tried and dropped. How do I stop that? · How do I keep a list of dead ends for my AI agents? · Can an AI agent know what was already tried in my project?

The work

When an approach fails, the agent records it with remember_failure, including why.

ArcRoom keeps
  • Saved decision
The next agent gets

Tried before and failed (Sep 28): WebSockets on Vercel functions.

Step by step

  1. When an approach fails, the agent records it with remember_failure, including why.
  2. When a later request matches, the failure is handed over with the reason.
  3. The agent picks another route, here polling or a hosted push service.

What the agent sees

Tried before and failed (Sep 28): WebSockets on Vercel functions.
Why: connections close when the function ends.

Compared with today's tools

A "tried and failed" list in CLAUDE.md

Add "WebSockets don't work on our serverless host" to a known-dead-ends section the agent loads every session.

Setup
5 min
Ongoing
write it up after each failure

ArcRoom

Every agent reads the team playbook before it works and writes to it when it finishes. Nothing to update by hand.

Setup
2 min (npx arcroom init)
Ongoing
None

Works if someone writes it; with remember_failure the agent records the failure and the reason at the moment it fails.

Where it stops

Only failures someone recorded are known. A silent failure teaches the playbook nothing.

Try it in your project

Works with the agents your team already uses, and any agent over MCP or the HTTP API.

More in solo builders