Herendil Advice and software for companies bringing AI agents into their own work

We help companies put AI to work on problems they already have. Most of the effort goes into the work itself: learning what the company does, finding the places where an agent can help, and building something small enough to trust.

Advice. We study how your team works today and recommend where agents and models belong, and where they would only add cost. The recommendation comes in writing, so you can act on it with or without us.

Building. We write the first working version with your engineers, from the agent itself down to the inference it runs on and the systems it connects to.

The company brain. Much of what a company knows is scattered across threads, tickets and the memory of a few people. We help gather it into a form that people and agents can both query, under the permissions it already has.

We work in short steps. Each one ends with something you can read or run, such as a written finding or a working prototype, and you can stop after any step and keep what was made. Four habits guide the work.

1. We begin with the work. Before we choose a model, we sit with the people who do the job and write down what they actually do, including the steps that live only in their heads.
2. We build narrow first. One agent that handles one task every day teaches more than a broad demo. We widen the scope once the narrow version has earned trust.
3. We test against your judgement. Evaluations use real cases from your team and are scored the way a colleague would review the work.
4. We leave you able to carry on. Your team owns the code, the prompts and the evaluations, along with a written account of how they fit together.

Most of this is ordinary engineering discipline applied to a new kind of software. Models change every few months. A clear record of the work, and tests built from it, carry over from one model to the next.

We write to work out what we think, and to give you a way to judge us before you call. The first essays are in preparation. Each is short, argues one point, and says where it could be wrong.

1. 1What a company brain is, and what it is notA working definition, and the three things it is most often mistaken for.
2. 2Agents need job descriptionsScope, inputs, and the person who checks the work.
3. 3Start with the records you already keepTickets, threads and meeting notes hold more than most knowledge bases.
4. 4Inference is a line itemWhat running models costs once the pilot ends, and how to plan for it.
5. 5Review an agent the way you review a colleagueEvaluations built from real cases, scored by the people who know the work.

If something here describes a problem you have, we would like to hear about it. A first call is a conversation about your work. We ask questions, and afterwards we send a short note on what we heard and where we would start.