Loading…
Systems that act.
From understanding to finished work, with a person able to check the result.
Agents
Sometimes the best interface isn't another interface.
We explore software that understands what someone wants and helps complete the work. Some of it is agents; much of it is plain automation and conventional software. Where one ends and the other begins is a design decision, and every result comes back in a form a person can check.
How we build them
01 — Workflow Agents
One task, start to finishAgents that gather context, plan the steps, use their tools and hand back a result.
02 — Automation
Routine work, handled reliablyWell-understood tasks, automated plainly, with judgement kept for the cases that need it.
03 — Experiments
Prototypes and open questionsThings we are trying, labelled with what we learned and what is still unknown.
04 — Conventional & Multi-tool Systems
Classic software and many-tool agentsDeterministic software where certainty matters, and agents that reach many tools where flexibility does.Understand. Then act.
01Understand
Work out what the person actually wants, and what is still missing.
02Plan
Break the job into steps and decide what to find out first.
03Use tools
Reach the maps, messages and data the job needs, and read the results.
04Verify
Check the work against something real before relying on it.
05Act
Carry out the work within limits agreed in advance.
06Report
Hand back the result with a record of what was done and why.
The intelligence behind it.
An agent is only as good as the context it is given and the checks it runs.
