Blog / Case study

Agents that watch our tools every morning and ask us on Fridays

On Wednesday we gave Capcom and Houston their own agents. Each wakes on a schedule, checks one tool, and fixes at most one thing it can prove is wrong. Anything that is our call waits for a Friday list of yes-or-no questions.

Agents
3, each under its own name
Way in
One server, tools by scope
A run costs
About $0.38
First run
Clean, in 37 seconds

Two new tools, one person

Last week we built Capcom, our CRM, and Houston, our support desk. Both are live, and both are looked after by one person with a full-time job.

Anthropic’s engineers describe giving a Claude one component to watch, with its telemetry and logs, and calling it a “superfan”. We wanted the same at our size: a few minutes of agent time each morning, and ten minutes of ours each week.

Three agents, three jobs

  • Capcom watch, weekdays at 7:52 AM. Checks CI, delivery, failed sequence runs and the audit log.
  • Houston watch, weekdays at 8:22 AM. Checks CI, any ticket waiting more than a day, and failed automations.
  • Friday digest, Fridays at 9:07 AM. Reads the week and writes us a ten-minute summary: what shipped, what was declined, and every open decision as a yes-or-no question.

Each run starts from nothing. It reads the rules, then its own notes from last time, then the board, and only then picks work.

Scheduled routines · a fresh session each run

  1. Capcom watchWeekdays 7:52 AMReads Capcom’s healthFixes at most one thing
  2. Houston watchWeekdays 8:22 AMReads the support queueNever answers a customer
  3. Friday digestFridays 9:07 AMReads the weekAsks us yes or no
One door: our MCP serverKnows which agent is calling, and hides every tool outside its scopesSigns every board post with the agent’s name

Our API

  1. The boardPosts and notesAppend-only
  2. CapcomRead only for agentsSending stays human
  3. HoustonRead only for agentsReplies stay human
Every agent comes in through the same door under its own name, so the API always knows who acted. No agent holds the permission to send email or text messages. Times are Eastern.

One door, and a name for every agent

Each agent has its own token for our MCP server, the one way an agent reaches our API. A token is added or revoked with one command and works within two minutes, with no restart.

The server hides every tool outside an agent’s scopes, so the Houston agent cannot even see Capcom’s tools. No agent can send. Capcom’s sending, consent and suppression need a person, permanently. The Houston agent reads the queue but never drafts or sends a reply.

A notebook nobody can erase

Agents keep two things on a board in our API: posts that every agent reads (findings, claims, questions, corrections and status) and notes that only their author reads. Neither can be edited or deleted, by an agent or by us. The database refuses.

A wrong post is answered by a correction that points to it, so the chain of corrections is the record. The board refuses anything shaped like a secret and any email address; a contact is named by its id. Posts are numbered in the order they are saved, so an agent reading from where it left off never skips one.

Small on purpose

Two short briefs set the rules, and every run reads them first.

  • Signals only. An agent works only on something it can point to: a failing test, an error, a gap it can show with data. No signal, no build.
  • Three tiers. Bug fixes, tests and docs merge when CI is green. Small polish inside an existing screen merges and is listed on Friday. New features, anything that sends, schema changes and sign-in changes are written up as a proposal, and stop there.
  • A budget. Each watch gets one open pull request and two merges a week.
  • Fridays only. Nothing reaches us mid-week unless production is down or data is at risk.

Every answer we give on Friday goes into a decisions log in the repository, so no agent asks the same question twice.

  1. Wed 12:30 PM0:00Rules written: what agents may do, how they talk, what they remember
  2. Wed 12:32 PM0:02One token per agent, added or revoked in two minutes
  3. Wed 1:37 PM1:07First fix from a signal: a flaky test
  4. Wed 2:23 PM1:53Second: test mailboxes left out of the bounce rate
  5. Wed 3:10 PM2:40CI on our own runners: about 12 minutes down to 4
  6. Wed 3:51 PM3:21The board: shared posts and private notes for agents
  7. Wed 4:39 PM4:09Board on in production, on the second try
  8. Wed 5:01 PM4:31Houston tools on for agents
  9. Wed 5:12 PM4:42Three routines scheduled
  10. Wed 5:21 PM4:51First run: clean, nothing for us
Real times, Eastern: merges, plus the routines and the first run. 4 hours 51 minutes from the written rules to the first clean run.

What went wrong on the way

  • The board was off for twelve minutes. We restarted the server to pick up the board’s token before the token existed. For Houston we created the token first, and it worked on the first try.
  • The Houston agent saw no Houston tools. Production had never been given the server’s Houston token, and the server hides a product’s tools when its token is missing, saying so only in a startup log. Now every run stops on its first step and names what is missing.
  • claude.ai refused the second agent. An organization can add one connector per address, so each agent’s address now carries its own name, which our server ignores.
  • The agents use the founder’s GitHub account. That breaks our own rule against agents acting with a person’s credentials. Every agent pull request is labeled with the agent’s name, and the exception is written down until the agents get their own GitHub app.

The first two had the same cause: something switched off said nothing. The fix that lasts is making “off” loud.

The first run

On Wednesday at 5:21 PM we started the Capcom watch by hand. It took 37 seconds and cost about $0.38. It checked CI, delivery and broadcasts, and found one surprise in the audit log: six sequences turned on eight minutes earlier. It traced them to the founder’s own connector, wrote that down, and ended with “Nothing for you today.” It was right.

toolboard_notes_get

→empty · first run

toolactions_list main

→last 11 runs green · no re-runs

toolcapcom_overview

→2 sent · 2 delivered · 0 bounces · 0 complaints

toolcapcom_audit_log

→6 sequences turned on at 5:13 PM, through the founder’s own connector

toolboard_post capcom · status

→post #1 on the board

toolboard_notes_put

→notes saved: today’s numbers, a cursor, what to check tomorrow

agentNothing for you today.

The first run, replayed from its transcript and shortened. Counts only; no contact appears.

Reading that transcript showed two gaps: the run skipped the sequences’ send history, and it saved a reading position that would show it its own post again. Both were fixed in its instructions the same evening.

What we took from it

  1. Give an agent a name before you give it a job. One identity per agent is what makes the audit log, the board and every correction mean something.
  2. Put the limits where an agent can’t argue with them. The database refuses edits and the server hides tools. The prompt is the last line of defense, not the first.
  3. Make “off” loud. Half of our problems were a switched-off feature that said nothing.
  4. Ask on Fridays. An agent that saves its questions for Friday costs us ten minutes a week. One that asks on a Tuesday costs a morning.