Assign roles across the proactive AI agent teams

In Days 3 to 7 you built separate proactive agents: intake, chronology, drafting and evidence-review. Each one watched its own part of the sample matter and acted when its trigger fired.

A Factory groups those agents into proactive AI agent teams. A team is a set of agents that share one responsibility. The Factory has three:

• Matter Intelligence builds and maintains the factual record.
• Work Product prepares deliverables from that record.
• Review checks the work and prepares the attorney queue.

Work moves in one direction, toward the attorney. Problems found by Review go back to whichever agent owns the faulty output.

The point is clear responsibility. Suppose a late email shows that goods were delivered on 14 March, not 4 March. With teams, you know exactly where that change belongs. It goes into the timeline, which belongs to Matter Intelligence. The drafts are then revised by Work Product, and Review checks that every deliverable now agrees. Without teams, three agents might each try to fix the date in their own way.

The Matter Intelligence team contains the agents that establish what the record supports. It does not draft or argue.

• Intake agent: keeps the document register current. When a new permitted document arrives, it adds an entry or raises a clear exception. It does not restart the whole task.
• Chronology agent: maintains the source-linked matter timeline. When an entry changes, it records why, and it keeps conflicting accounts visible instead of choosing one.
• Evidence-review agent: keeps the evidence index and the list of evidence questions. Each question is tied to an issue and to the material that prompted it. It also prepares a bounded discovery packet for the attorney to consider.

Example: a second batch of three permitted documents arrives. The intake agent registers all three. The chronology agent adds two new events with source links. The evidence-review agent notes that one invoice refers to a purchase order that is not in the file.

None of these agents decides what that gap means legally. Each one only makes the gap visible and traceable.

The Work Product team holds the drafting agents. They prepare deliverables such as:

• the matter brief
• the issue matrix
• the negotiation brief
• client explanation drafts

A practical design is one drafting agent per deliverable. That way each draft has a single owner, and when a draft is wrong you know which agent to inspect.

Drafting agents read the factual record. They do not write to it. Suppose a draft needs a fact the record does not contain. The agent marks the gap and routes the claim to the verification queue, as you practised on Day 5. It does not supply the fact from general knowledge.

Example: the negotiation brief agent is summarising the counterparty's position on late-delivery penalties. The only source is a counterparty email that asserts the goods were late. The agent writes "the counterparty states," links the email, and keeps the allegation from becoming an established fact.

Every output from this team is a draft. An attorney remains responsible for whether and how it is used.

The Review team checks the work of the other two teams. A typical set of agents:

• Source-link checker: confirms that each claim points to a source that actually says it.
• Verification queue agent: runs the verification queue, the list of claims that need review. Examples are an allegation stated as fact, a date with no source, or a legal authority the attorney has not yet verified.
• Consistency agent: compares deliverables against each other and against the record.
• Attorney queue agent: assembles the attorney queue, the list of items waiting for an attorney decision, each with its sources.

Review reports problems. It does not rewrite drafts or edit the record. It sends each finding back to the agent that owns the output.

Example: the consistency agent sees that the matter brief gives 4 March as the delivery date and the issue matrix gives 14 March. It does not pick one. It sends the conflict to both drafting agents and to the chronology agent. If the record itself cannot settle the date, it lists the question in the attorney queue.

Each agent in every team keeps a job definition: a short written record of five things.

1. Trigger: what makes the agent act.
2. Inputs: what it reads.
3. Outputs: what it writes, and where.
4. Limits: what it must not do.
5. Stop rules: the conditions under which it halts and raises an exception instead of continuing.

Here is the chronology agent's definition for the sample matter:

• Trigger: an accepted change in the document register.
• Inputs: the new register entry, its source document and the current timeline.
• Outputs: added or changed timeline entries, each with a source link and a reason for the change.
• Limits: writes only to the timeline. Does not judge whose account is more credible. Keeps both accounts when they conflict.
• Stop rules: stop and raise an exception if the source is unreadable, falls outside permitted access, or cannot be placed in time. Also stop when the attorney has placed a hold.

Written definitions let you compare agents side by side. That comparison is how you will spot overlaps in the next step.

Give each section of the record exactly one owning agent. The owner is the only agent that writes to that section. Other agents may read it and may send proposed changes, but only the owner applies them.

An overlap is any output that two agents could change. Here is a common one. The chronology agent writes event dates. You might also let the evidence-review agent "correct" a date when it finds a better document. Now two agents write to the timeline. Whichever writes last wins, and the reason for the earlier entry can be lost. A newer document can also silently replace a better-supported one.

The fix is that the evidence-review agent sends the proposed date, with its source, to the chronology agent. The chronology agent records the change and the reason.

To find overlaps, list every output and, next to it, every agent whose job definition says it writes there. Any output with more than one writer is an overlap to remove. Use the sliders to see how quickly the risk grows when several agents share write access.

Some decisions belong to the attorney, whatever the Factory's design. No team, and no agent in any team, may:

• release work outside the Factory
• decide privilege
• make discovery requests
• approve advice, communications or filings

The attorney also verifies legal authority in the appropriate source before a draft that relies on it is used.

Write these as limits in every job definition, not just once for the whole Factory. That way each agent's own definition tells it where to stop. Teams prepare decisions; the attorney makes them.

Example: after a cycle, the attorney queue holds three items:

• a client explanation draft that Review found consistent with the record
• a document carrying a "confidential" header that may be privileged
• three evidence questions in the discovery packet

Even though Review found no problems with the client explanation, it is still not sent. The Factory flags the possibly privileged document but does not classify it. The discovery questions stay questions until the attorney decides what, if anything, to request.