Agents that hold up
on real data.

What we mean by AI

Most useful AI work in a business is not a chatbot. It is a task someone does every day, by hand, that follows rules most of the time and needs judgement some of the time: triaging requests, reconciling records, extracting fields from documents, checking that something matches something else.

A language model can do the judgement part reasonably well. Everything around it, getting the right data in, taking the right action out, and knowing when it got it wrong, is ordinary engineering. That engineering is most of the work, and it is where projects succeed or fail.

Agents are programs that choose tool calls

An agent is a loop: the model reads the task and the context, decides which tool to call with which arguments, sees the result, and decides again until it is done or hands off to a person.

The tools are the important part. Each one is a typed function with a name, a description the model reads, and a schema for its arguments:

{
  "name": "issue_refund",
  "description": "Refund an order. Only for orders delivered in the last 30 days.",
  "input_schema": {
    "type": "object",
    "properties": {
      "order_id": { "type": "string" },
      "amount_cents": { "type": "integer", "minimum": 1 }
    },
    "required": ["order_id", "amount_cents"]
  }
}

Well-defined tools make agents predictable. Narrow tools that do one thing, validate their own inputs and are safe to retry are far easier to trust than one general tool that can do anything.

Retrieval: give the model the right page

When answers depend on your documents, policies or records, the model needs the relevant passage in front of it. Retrieval pipelines split documents into chunks, index them, and fetch the ones most relevant to each request.

Retrieval quality is measurable on its own: for a set of real questions, does the right passage come back? If it does not, no prompt will save the answer. Measuring retrieval separately from generation tells you which half to fix.

Evaluation against a threshold

The difference between a demo and a system is evidence. We build an evaluation set from real cases, including the awkward ones, write down the expected outcome for each, and agree a pass rate before building. Exact checks come first (right tool, right arguments, right refusal); free-text outputs are graded against a rubric, with any model-based grader calibrated against human judgement.

The evaluation runs on every change to prompts, tools or model. If the score falls below the threshold, the change does not ship. Evaluate before you ship goes through this step by step.

Running it where your data lives

An agent in production needs the same things as any other service, and a few more:

• Your infrastructure, so data stays where it already is and access follows your existing controls.
• Retries and idempotency, so a transient error cannot cause an action twice.
• A dead-letter path for cases that still fail, and alerts to a channel someone reads.
• A person on exceptions, with the agent's reasoning attached to what it routes.
• Logs of every tool call, so any action can be traced back to the input that caused it.

How Rump Labs works in AI

We build tool-calling agents, retrieval pipelines, evaluation harnesses and workflow automation on the client's infrastructure, measured against a threshold the client sets.

The usual starting point is One Workflow, Automated: one task your team repeats daily, mapped with the person who does it now, built and deployed in ten business days, connected to up to three systems you already use, with retries, a dead-letter path, alerts and a runbook. If it does not run end to end on your data by day ten, you do not pay the second half.

We also teach it. Our hands-on workshops end with one working agent per person, as at the Heritage Institute of Technology, where third-year computer science students spent an afternoon building their first AI agent.

Where it shows up

Where agents do the repetitive part of the work, with a person on the exceptions.

Where to start

Further reading