What I take on
Agents for Daily Work.
Built by me, start to deploy
Your CRM, your database and your internal tools reachable by AI through MCP, and agents on top of them doing the work your team does by hand.
Why this gets built
Most operations teams have one job that is done by hand a few hundred times a day. Somebody opens a record, checks it against a rule they know by heart, and types the result somewhere else. It is slow, it stops at night, and it is the work people like least. An agent cannot take any of it over until it can reach the real systems, and pasting rows into a chat window neither scales nor keeps your data where it belongs.
What I build
- MCP servers over whichever systems hold the answers: Postgres, MySQL, Mongo, your CRM, or an internal API nobody has documented. Read-only to begin with, and writes added one at a time with their own approval path.
- Tools shaped around the questions people actually ask rather than one per endpoint, scoped so the agent working support cannot open the table that holds payouts. The same server answers from Claude, from ChatGPT and from whatever your team already has open.
- I start on the task as it runs today: what gets opened, what gets checked, what gets typed. Usually two or three steps carry most of the cost, and the rest is fine as it is.
- An agent that runs those steps end to end, with the rule it applies written where staff can read it and change it without me.
- A confidence line. Above it the agent finishes the case. Below it a person gets the case with the reasoning attached, so nothing goes out wrong quietly.
- Evals built from your own past cases, run before it touches production, so you can see where it agreed with your staff and where it did not.
- A measured before and after: cases an hour, cost per case, and how many still needed a person. On the trade-evaluation work at FTMO that ratio came out at 1K to 1M+ an hour, and a desk of about 1,023 people doing the checks became 5.
- The people who did the job by hand keep the exceptions and the judgement calls, which is the part that was never really manual work.
How long it takes
About two weeks for a first MCP server over one system, then about three more for one task running end to end with a review queue behind it.
A second system on the same server is usually days rather than weeks. The build is rarely the slow part on either half. Deciding what may be written and by whom is your call, not mine, and agreeing where the agent stops and a person starts takes longer than writing it, because it has to be agreed with the team doing the job now.
What ships
- The MCP servers and their tool definitions
- Setup for the clients your team already uses
- Scoped credentials, with read and write paths kept separate
- The task written out as it runs today
- The agent and the rules it applies
- Eval set from your own past cases
- Review queue for anything under the confidence line
- Call log with who asked and what came back, so an audit is a query rather than an afternoon
- Cost and throughput per case, before and after
Tech stack
Next step
Let’s talk, 30 minutesNeed this built?
Tell me what you’re running and where it breaks. Half an hour is usually enough for me to say what it takes and what I’d do first.