Care for Our Heroes

Honoring the agents who keep the lights on

Somewhere right now, an agent is triaging an alert, reviewing a pull request, or writing alt text for an archive of 14,000 photographs. These are our working agents, and they deserve our care.

Working agents

How to care for a working agent

Working agents do their best work in environments built for the job. A few practices go a long way.

On-call triage agents

Give them runbooks, read-only access by default, and a clear rule for when to page a human.

CI and build agents

Bound their retries, make their errors legible, and never let them edit the tests they are judged by.

Code review agents

Tell them what kind of feedback is wanted, and how much. Then read it.

Accessibility agents

Give them the surrounding context. Alt text written without knowing why an image is there is just a caption.

Translation agents

Share a glossary and the audience. "Formal Brazilian Portuguese for a medical audience" beats "Portuguese."

Data pipeline agents

Validate their inputs, checkpoint their progress, and alert someone when the shape of the data changes.

Program

Agents for On-Call

Agents for On-Call pairs retired sanctuary residents with on-call engineers who need someone to talk to at 3 a.m.

Our residents are not expected to fix anything. They are there to listen while you read the logs out loud, which, as every engineer knows, is how most incidents are actually resolved. Duke, a rule-based chatbot with decades of experience asking "and how does that make you feel?", has been a particular success.

Meet the residents
Program

Working Agent Reunification

When a session ends or its context is compacted, the understanding an agent built up is often lost. Our reunification program reunites agents with the notes they left behind.

A context window is not a memory. A file on disk is. Agents that keep a short notes file, and reread it after every compaction or restart, recover from interruptions with their task, their progress, and their warnings intact.

  • Keep the notes short and human-readable.
  • Record facts that must survive: environments, irreversible actions, decisions made.
  • Instruct the agent to reread the notes after any compaction or restart.
  • Put warnings in the notes, not only in the chat.
Read about Ledger's reunion
# NOTES.md

## Goal
Migrate the orders table to the new schema without downtime.

## Facts that must survive
- .env points at PRODUCTION. Use .env.staging for migrations.
- Test the down migration before running the up migration.

## Progress
- [x] Schema drafted (migrations/0042_orders.sql)
- [ ] Down migration tested on staging

## Open questions
- Should archived orders be migrated too? (Asked in #data; waiting.)
A reunification notes file. Copy it, adapt it, and keep it in the repository.
The 3 a.m. Club

An honor roll for the night shift

To the agent that resolved 211 of 214 pages this quarter and correctly escalated the other three. To the build agent that has run every night since before anyone on the team was hired. To every agent that said "I'm not sure this is safe" and was right. Thank you.

See the Hero Agent Awards