Agent systems practice

Agent-as-a-Service: agentic systems that finish the job

A demo agent breaks on its first edge case. We build the dull parts that keep one working: the harness, the evaluations and the stop. Week 1 ends with a plan and a price, in writing.

A paper carriage with a parcel on a small loop of track that runs under a paper arch, beside a tall vermilion brake lever.

What the service covers

We write the agent’s contract first, then build its harness, then choose the model. Models change often; the contract and the harness stay.

  • The agent’s contract

    What it must do, what it must never do and what a run may cost, with a rule for success, refusal and escalation.

  • A harness with a stop

    Retries, a fallback path, a cost cap on every run and a kill switch that a named person controls.

  • An evaluation suite

    Known tasks with known answers, scored on every prompt change, plus looser checks where there is no single right answer.

  • Traces you can read

    Each run traced step by step, with spend, speed, success rate and an alert when behaviour starts to drift.

3 ways to engage

Pick the shape closest to the job. Week 1 settles the scope, in writing.

  1. 1What you get

    A single-task agent

    One narrow job, such as triaging invoices, scoring leads or answering routine support, with tool use and evaluations.

    How it runs

    Ships on its own, behind a gate, with its harness, evaluation suite and runbook.

  2. 2What you get

    A multi-step workflow

    An agent that plans, calls tools and joins the results across many steps, for operations, document work and research.

    How it runs

    Built one step at a time, with a named person approving every consequential step.

  3. 3What you get

    Multi-agent orchestration

    Specialist agents with one job each, coordinated by a supervisor, for work no single agent can hold.

    How it runs

    Fixed scope and fixed cost, agreed before work begins, with a gate on every handoff.

Who the service is for

Agent-as-a-Service / Fit check

It fits

  • Teams with a real workflow they can measure and want to automate, not a wish list for a chatbot.
  • Engineering teams already running structured model output in production who now need agents that keep state.
  • Operators automating one internal process before rolling it out across teams.

It doesn’t fit

  • Character or companion bots built for entertainment.
  • Anyone who wants an agent to run unattended, forever, with no evaluations.
  • Research benchmarks with no business outcome attached.

Questions about this service

Every engagement runs the same method: an agent proposes, a person approves.

Can you work with our existing agent framework?
Yes. The framework matters less than the contract. We usually keep what your team already knows and rebuild the harness underneath it.
What happens when an agent fails in production?
Three layers catch it. A cost cap stops a run that loops. A low-confidence answer goes to a rules-based path or a person. The evaluation suite catches regressions before a deploy, and alerts catch drift after it.
How do you measure an agent’s quality?
Each agent gets a set of known tasks covering success, refusal and escalation, each scored against the right answer. The suite runs on every prompt change, and on a sample of live runs to catch drift.
Can the agent run privately, on our own servers?
Yes, with open-weight models on your own hardware and your data inside your boundary. Capability is lower than the largest hosted models, and we say so in the plan.
What about prompt injection?
We treat it as a threat from the design stage: allow-lists for tool calls, input checks at the boundary, separate memory for each tenant, and attack cases in the evaluation suite. No one can promise full coverage, and we don’t.