What we do
Four stages. One promise.
Agents that know your business, can be trusted with it, and can be found by it. We do the stages in order, and you can stop after any one and keep everything.
Stage one
Audit
Two weeks
Know which of your tasks an agent can actually do, with evidence from your own work.
- Who it's for
- A team of five to fifty that tried an AI tool and quietly went back to doing the work by hand. Or one that hasn't tried, and wants to know where to start without buying a platform first.
- What we need
About four hours of your people's time across the two weeks: two interviews with the people who actually do the work.
Ten real examples of each task we look at: the emails, quotes, tickets, reports, whatever the task produces. No access to your systems.
- What you receive
-
task-map.mdEvery recurring task we found, with one of three verdicts: automate, assist, or keep human. The reason in one line each.constraints.mdThe rules your people apply without thinking, written down for the first time. This file is the seed of every skill that comes after.trial/One task we ran an agent on against your real examples, outputs side by side with what your people produced. The verdicts are evidence, not opinion.readoutForty-five minutes with whoever decides. What to hand over first, what to leave alone, and why.
- What's next
- Nothing, if the map says so. We will tell you when an agent is not the answer. Otherwise a Foundry engagement scoped to the top rows of the map.
Stage two
Foundry
Three to five weeks per department
Your knowledge, written as skills your agents follow and your people can check.
- Who it's for
- A team whose audit found tasks worth handing over. Or anyone already running Claude, Codex or a similar agent and getting generic output from it, because the agent has never been told how this company works.
- What we need
Your best person on each task, for about two hours a week. They are the author; we are the editors.
The examples from the audit, plus the ones they would be embarrassed by. Access to the agent product you use, or we recommend one.
- What you receive
-
skills/<task>/SKILL.mdOne skill per task: the rules, the steps, the checks, and the reference files the agent reads. Readable by a person, followable by an agent, versioned in your repository.skills/<task>/evals/Twenty real cases per skill with the output your expert would accept, and a script that scores a model against them. When the model changes, you know in ten minutes whether the skill still holds.installThe skills running inside your agent product, in your account, with every output it produced during the build kept for you to read.handoverYour team changes a rule, re-runs the evals, and ships the change, with us watching. If they can't do that, we are not finished.
- What's next
- A monthly care option: on every model release we re-run your evals and fix what broke. Or you do it yourself. The scripts are yours either way.
Stage three
Review
One to two weeks
Decide what the agent may decide, and make the rest wait for a person with the reason attached.
- Who it's for
- A team whose agents now produce more than a person can read. Or one whose people read everything, because nobody ever said what could be trusted, so the agent saves no time at all.
- What we need
The skills from Foundry, or your own. One week of real agent output. The person who currently approves things, for an hour.
- What you receive
-
review-policy.mdFor each task: what goes out unread, what waits for a person, what is never sent. Written as rules the agent applies to itself.held queueDrafts that hit a rule wait with the reason attached, in the tool your team already uses: email, Slack, or the agent product's own queue.weekly digestWhat went out, what was held, what a person changed. The rules get tightened or loosened from that evidence, not from nerves.
- What's next
- The policy becomes part of the skills, so it travels with them when the model, the tool, or the team changes.
Stage four
Readable
Two weeks
Be found, compared and booked by other people's agents, not only by people.
- Who it's for
- Any business that gets customers through search, directories or referrals, and wants the same when the customer sends an agent instead of coming themselves. That is starting to happen now, and most websites are unreadable to an agent beyond a phone number.
- What we need
Your real offer: services, area, hours, the booking path, the policies you hold to. Your website's repository or hosting, or we hand you the files and instructions.
- What you receive
-
/.well-known/offerA structured, versioned description of what you do, where, when, and under what rules, at a path an agent can find without guessing from a photo.structured dataThe same facts marked up on the pages you already have, so search and agents read one consistent story./bookA booking path an agent can complete: a form or an endpoint that accepts a request, confirms it, and hands it to a person.monthly checkWe send an agent to find you the way a customer's would, and report what it saw and what it could not do.
- What's next
- As agent shopping grows, the offer file grows with it. The format is open and lives on your domain. You are not locked to us.
What we don't do.
- Generic chatbots or template automations.The floor for those rises every month and you can get them anywhere. They encode the task, not the judgement, which is why they go back to being done by hand.
- Train or fine-tune models.The models are good. What is missing is what they know about you, and that is a writing problem, not a training problem.
- Replace the people who hold the knowledge.We write down what they know so an agent can follow it. The judgement stays with them, and Review keeps it there.
- Invent numbers.No efficiency percentages, no return-on-investment calculators. The eval set and the side-by-side outputs are the evidence.
- Lock you in.Skills are files in your repository. Evals are scripts. The offer file is a public format. If we vanish, everything keeps working.
Start with the audit.
Thirty minutes to find out whether it fits. If an agent is not the answer for your team, we will say so on the call.