# PDR 000: Start with advisory matching for ready Work Graph tasks

- HTML version: https://robbiepalmer.me/projects/agent-coordinator/pdrs/000-advisory-work-graph-routing
- Project: Agent Coordinator
- Status: Superseded
- Opened: 2026-09-04
- Decided: 2026-09-04
- Superseded by: https://robbiepalmer.me/projects/agent-coordinator/pdrs/001-severe-vulnerability-resolution.md

## Context

Work Graph records which tasks are ready, how they rank, what blocks them, and
what evidence can prove completion. It does not know which person or agent has
the capability, authority, budget, and current capacity to do a task well.

Agent capacity changes as subscription limits reset, API budgets move, and
different models become available. Treating that capacity as a reason to launch
work can fill a human review queue with weak output. Putting provider and model
details in Work Graph would also tie a durable plan to short-lived execution
choices.

The first Agent Coordinator workflow needed to test whether matching could move
ready work forward without taking ownership of the plan or automating choices
that still required human judgment.

## Decision

Start with advisory matching of ready Work Graph tasks.

The coordinator will read ready work and owner policy, compare task constraints
with an actor's declared capabilities and observed outcomes, and propose a task,
actor, and bounded context package before anyone claims the lease. The owner
will see the proposed match for the first workflow.

Work Graph remains authoritative for priority, dependencies, readiness, leases,
attention, and completion evidence. The coordinator may claim and renew a lease
for the selected actor, but it may not change the plan, resolve an attention
request, or declare completion without the evidence required by Work Graph.

Automate matching only inside owner-defined limits for task class, authority,
access, budget, and concurrency. If no eligible actor can produce a useful
outcome, leave the capacity idle. Prepaid capacity may break a tie between
otherwise suitable workers, but token use is not a product outcome.

People using their own agents retain control of those agents and their
credentials. A later contributor workflow may offer bounded work to a
contributor-owned agent, but the coordinator will not collect or intermediate
the contributor's subscription credentials.

## Consequences

The first workflow keeps a person in the routing loop, so it cannot allocate
capacity continuously. It can still test the matching inputs, context package,
lease handoff, and evidence record before those choices gain unattended
authority.

Provider names, subscription allowances, and model prices stay out of the
durable work plan. The coordinator needs a separate current view of actors and
capacity, and it must learn from completed, abandoned, and reassigned work
instead of trusting model labels.

Some paid capacity will expire unused. That is preferable to assigning work
whose outcome lacks value or still needs open-ended human review.

## Evaluation

The workflow succeeds when ready work spends less time waiting, matched actors
finish the tasks they receive, and a session handoff preserves useful progress.
It fails when it keeps cheap agents busy on low-value work, prizes subscription
use over project outcomes, or transfers the real evaluation burden to a human.

## Migration note

This record preserves the first-release policy published in the original
project narrative on 2026-09-04 and the product constraints later mixed into
ADR 000. The migration changes the record type, not the decision, date, or
status.

## Related records

- Project: [Agent Coordinator](https://robbiepalmer.me/projects/agent-coordinator.md)
- Implemented by ADR: [ADR 000: Use provider-supported authentication for routed sessions](https://robbiepalmer.me/projects/agent-coordinator/adrs/000-provider-supported-session-routing.md)

---

Markdown index of this site: https://robbiepalmer.me/llms.txt