Tools · 2026

Library of Agents

A shelf of AI agents for a wide range of industries. Pick one up and put it to work.

Status
Shipped
The Library of Agents home screen: the headline "A shelf of AI agents for a wide range of industries", industry filters, four featured agents (Research, Invoice, Clinical Notes and Property Research) each with a status tag, and colourful tiles to browse by industry.

People often start AI work with a blank chat. Every time, they have to work out the role, the context, the inputs and what good looks like.

A library gives them a starting point. Pick an agent built for one specific job, see what it needs and what it can’t do, then put it to work.

It’s as much a service and governance problem as a prompt problem. Discovery, inputs, review, feedback and maintenance all shape the experience. A trustworthy library also helps you decide when not to use an agent.

  1. Problem

    A blank chat makes people reinvent the role, the context and the expected output every time. A library can make good starting points easy to find, while keeping the person in control.

  2. Research

    Planned: test agents on representative tasks with a rubric for relevance and completeness, including missing-context and adversarial cases.

  3. Prototype

    The flow: open an agent card, see its purpose and limits, add context in a short form, review the result, then copy or refine it. [Verify: what's built today]

  4. In people’s hands

    Next: put it in front of designers and knowledge workers with repeatable, bounded tasks.

  5. What happened

    My bet so far: narrow, transparent agents beat a big catalogue of overlapping ones. Broad agents with vague promises are hard to evaluate.

How I’m approaching it

The Double Diamond, as planned. Plans loop back once real people get involved, so this one does too.

  1. Discover

    Find the tasks people repeat often enough to deserve an agent.

  2. Define

    Organise the catalogue around what people are trying to get done.

  3. Develop

    Prototype the agent cards and one working agent.

  4. Deliver

    Test output quality, and especially the failure cases.

  5. Define

    Narrow the scope wherever agents start to overlap.

  6. Develop

    Build an explicit model for evaluating and maintaining agents.

UX practices

  • FindabilityAgents are organised by the job someone needs done, not by technical category.
  • Expectation settingEach agent shows its purpose, inputs, example outputs, limits and when a human should review.
  • Feedback and recoveryEasy to edit context, retry safely and flag a poor result.

Service design

  • Service blueprintingThe browse-and-run interface connects to orchestration, model providers, logging and maintenance.
  • AI governanceClear ownership, review criteria, versioning and escalation for harmful or low-quality output.

Privacy notice