A short lexicon of applied AI agent – inference

Herendil

/ˈher·ən·dɪl/ n.

  1. A small company that helps other companies put AI to work on their own problems. We build agents, set up inference and connect what a company knows, so people and agents can both use it.
  2. Consulting and software for teams who want the work done inside their company and handed to them once it runs.

Bring in Herendil for the first agent, then run the next ones yourselves.

Book a call →

Entries agent call company brain context evaluation handover inference

agent

/ˈeɪ·dʒənt/ n.

  1. Software that is given a goal and works toward it in steps, using only the tools it has been allowed to use.
  2. In a company, a colleague with a narrow job and one owner who answers for it. It keeps a record of everything it did.

The agent drafts the refund and a person approves it.

See also context evaluation handover

call

/kɔːl/ n.

  1. A conversation, usually half an hour, about one piece of work in your company.
book a call phr.
To pick a time to talk it through with Herendil. Describe the problem as you would to a colleague; you don’t need a brief or a deck. By the end you will know what we would try first, and whether we are the right people to do it.
Find a time →

See also Herendil handover

company brain

/ˈkʌm·pə·ni breɪn/ n.

  1. What a company knows, decides and does, kept where people and agents can both reach it.
  2. The memory behind a useful agent. It holds the documents, and also each decision with its reasons and the person who made it.

Ask the company brain why the pricing changed in March.

See also context agent

context

/ˈkɒn·tɛkst/ n.

  1. Everything a model sees when it answers: the instructions, the question and any documents fetched for it.
  2. The usual cause of a bad answer. When context is missing or out of date, a larger model rarely helps. Fetching the right documents does.

Give it the contract and last week’s thread as context, then ask.

See also company brain inference

evaluation

/ɪˌvæl·juˈeɪ·ʃən/ n.

  1. A set of real tasks with known good outcomes, run against an agent before and after every change.
  2. How you know an agent works. Without one, every demo looks fine and every regression comes as a surprise.

The new model scored lower on our evaluation, so we kept the old one.

See also agent handover

handover

/ˈhænd·əʊ·və/ n.

  1. The point where a system we built becomes yours, with its code, its evaluations and the know-how to run and change it.
  2. Planned from the first week of every Herendil project. We judge our work by how well your team runs it without us.

Plan the handover before writing the first line of code.

See also evaluation call

inference

/ˈɪn·fər·əns/ n.

  1. Running a trained model to get an answer. It is the part of AI you pay for every time someone uses it.
  2. A set of choices about which model, where it runs, how fast it has to be and what each answer may cost. A small model on your own servers is sometimes the better one.

Move routine questions to cheaper inference and keep the large model for hard cases.

See also context agent