AGENT TEAM · SPECIALISTS AND COVERAGE
Agent team: independent specialists and a final check
Have legal, finance and security specialists review an eight-clause contract, and see what shared notes, voting and a final coverage check change.
Your experiment
Eight clauses, five with real issues, one of them in privacy, a domain nobody on the team owns. The specialists review in turn, reading each other’s notes, and any single flag blocks a clause. Count the issues stopped, then make the reviews independent, add the coverage check and try a majority vote.
Play replays the computed run over sixty inspection seconds; Next event steps, Restart rewinds, and Inspection time scrubs. Any input change recomputes the run and rewinds it without autoplay.
Event trace
Computed data
Final metrics
Deterministic team: what each specialist flags when reviewing alone is fixed (each finds the issues in its own domain except one blind spot, and legal also catches the currency clause C6); in shared-notes mode a specialist defers to an approval already given by the clause’s domain owner. No model is called and the clauses are invented.
Specialists, domains and one gap
Three specialist agents review an eight-clause supplier contract. Five clauses have real issues: unlimited liability (C1, legal), 120-day payment terms (C2, finance), no breach notification duty (C3, security), data stored in any country (C5, privacy) and fees converted at the supplier’s own rate (C6, finance). Reviewing alone, each specialist flags the issues in its own domain except one blind spot: finance misses C6, which legal does catch. Nobody on the team owns privacy, so C5 has no natural reviewer.
The opening run: shared notes cost a catch
The specialists review in order (finance, security, legal), each reading the notes of those before, and any single flag blocks a clause. Finance flags C2 and approves C6; security flags C3; legal flags C1 but, seeing that the finance owner already approved C6, defers to that approval. With no final check, C5 and C6 are approved: 3 of 5 issues stopped. Set How specialists review to Independently, without seeing others: legal’s view of C6 is no longer anchored, and 4 of 5 are stopped.
Failure boundary: coverage and voting
Set Final check to “Every clause needs its domain owner’s sign-off”: C5 has no privacy owner on the team, so it is escalated instead of approved, and the independent team stops all 5 issues. With shared notes the coverage check still misses C6, because finance did sign off on it. Change the Blocking rule to Two flags needed: each issue is seen by only one specialist, so nothing is blocked and 0 of 5 are stopped (1 with the coverage check). A majority vote suits redundant reviewers of the same question, not specialists of different ones.
Removing a specialist
Clear Security specialist on the team: with independent reviews and the coverage check, C3 is escalated because it lacks a security sign-off, and so is the clean clause C7, so all 5 issues are still stopped at the cost of one unnecessary escalation. Clear Legal specialist on the team without a final check: C1 and C6 are lost along with C5, leaving 2 of 5. Clear Finance specialist on the team in shared-notes mode: legal no longer defers on C6 and catches it, but C2 and C5 are missed; adding the coverage check escalates C2, C5 and the clean C8.
Reading the run and its limits
Play replays the brief, each specialist’s review, aggregation, the final check and the outcome over sixty inspection seconds; Next event, Restart, Speed and Inspection time control the replay, and any input change recomputes it. The table lists each clause, its domain, whether it has a real issue, who flagged it and the decision. The specialists are fixed scripts with stated blind spots and a stated deferral rule, not language models; real agents vary, which is why independence and an explicit coverage check are design choices rather than details.