All work
Lively & ARC IP Law · AI systems

Five agents with real boundaries

Most teams end up with one general-purpose assistant that is vaguely useful at everything. I built five that are each unambiguously responsible for something — and told them, in writing, what to refuse.

RoleDesigner & builder
Agents5
OrganisationsLively, ARC IP Law
DisciplineAI systems
Illustrative agent system showing the Lively and ARC IP Law rosters, routing rules and the due diligence path
Illustrative interface — no client or counterparty information

The problem with one assistant

A single general-purpose agent gives you confident answers in domains it has no business answering. Ask it to price an offer and it will cheerfully produce a number with no grasp of margin. Ask it whether a prospective client is safe to take on and it will tell you, without knowing that it should not be the one deciding.

The failure is not capability. It is that nothing owns anything, so nothing can be relied upon — which is the same failure as an operations team with no named owners.

Lively: a strategy-and-execution team

Three agents whose domains genuinely overlap, which is why the routing rules are the real design. A user's question rarely announces which agent it belongs to, so the boundaries are written down as a shared vocabulary, resolved once:

Underneath them sits a shared memory: a standardised knowledge base per client brand — brand, audience, voice, offers, channels, campaigns, performance, competitors, economics, pipeline, sales process, decisions and learnings. Fourteen files, the same shape for every account, so someone who has never touched an account gets a real answer in one prompt rather than a plausible one. Prompts are easy to copy. Accumulated client intelligence in a consistent shape is not.

ARC IP Law: research agents

Diligence for a law firm is the clearest case for hard boundaries, because a wrong answer is not merely unhelpful. Tony is built around four rules:

The detailed procedures load only when a task needs them, so the core instructions stay short enough to follow reliably.

The principle both share

Each agent is told what to decline, not only what to do. Jarvis executes; he does not set strategy. Bruce does not touch pricing. Tony researches and recommends; he never clears conflicts or accepts a client.

That refusal is what makes the answers trustworthy. An agent that answers everything tells you nothing about whether this particular answer is any good — and in a law firm, an agent that oversteps can create a problem a person then has to unwind.

How it gets maintained

Each agent has one canonical definition, versioned in a single place, with build scripts that generate the paste-ready instructions and packaged skill bundles from it. Updating an agent is one edit and a rebuild, rather than a message asking everyone to re-copy something.

Which is the same idea as every other system here: write it down once, give it an owner, and make the next use a configuration rather than a rebuild.

The interface above is an illustration built for this portfolio. No client, counterparty or knowledge-base content from either organisation appears on this site.

Next case The Final 6-Week Challenge