Agent reliability
Most agent demos stop at "it worked once." This series is the other half: permissions, undo, traces, handoffs, evals — the reliability work you do before autonomy is safe to ship.
7 posts · oldest first
- 2026.066 min
Agents Fail Quietly
A passing test and a working workflow are not the same thing. The gap between them is where production breaks.
- 2026.076 min
Give Your Agent an Undo Button
You can bound what an agent touches and measure when it fails. Neither one saves you if the mistake it makes cannot be taken back.
- 2026.076 min
Teach Your Agent to Ask for Help
Full autonomy is the wrong goal. The best agents know exactly when to stop and hand the decision back to you.
- 2026.076 min
Trust Comes From the Trace
You bounded what your agent can touch, gave it an undo button, and taught it when to ask. But when it does something surprising, can you reconstruct exactly why? That is the observability gap.
- 2026.077 min
Your Second Agent Is the Hard One
Going from one agent to many does not multiply what you can do. It moves every failure into the space between them.
- 2026.087 min
The Handoff Is Where Agents Break
You debug the agent that produced the bad output. The bug was in the message it got handed.
- 2026.088 min
Your Eval Suite Measures the Wrong Thing
Nearly a quarter of what goes wrong in multi-agent systems is a verification failure. The most common one is a check that ran, and said yes.