# Agent Coordinator

> A matching and handoff system that offers ready work to suitable people and agents without owning the work plan

- HTML version: https://robbiepalmer.me/projects/agent-coordinator
- Status: idea
- Started: 2026-09-04

# Vision

Keep available agents working on the most useful tasks they can perform, while
letting the Work Graph remain independent of models, subscriptions, machines,
and agent products.

The coordinator answers a different question from the graph. The Work Graph
says what is ready and what constraints apply. The coordinator finds an actor
whose capabilities, interests, authority, budget, and current capacity fit the
task.

# Why this is separate

Model choice changes quickly. Some work needs a frontier model for research or
architecture. Some can run on a cheaper model. Mechanical maintenance may be
suited to a free model. Available capacity also shifts as subscriptions reach
usage limits or a person loses focus.

Putting those details in the Work Graph would couple durable project knowledge
to a temporary market. The graph should keep task complexity, required
capabilities, authority, and evidence. The coordinator should map those needs
to the actors available now.

The initial routing model can use broad bands. Deep research and architecture
need frontier reasoning. Well-bounded implementation can use mid-priced models.
Mechanical maintenance may fit cheap or free models. The named models will
rotate too quickly to make them part of the durable schema. Evaluation history,
observed task outcomes, current prices, subscription allowances, and available
capacity should drive the mapping.

# Actors

An actor may be a person, an agent directed by that person, or an agent running
under authority owned by the project. Each actor brings claims about capability,
interests, cost, access, and availability. The coordinator must treat those
claims as inputs to test, not unquestioned facts.

For public contribution, people cannot hand a project access to a subsidized
agent subscription. That would break product terms and create a credential
nightmare. Instead, a contributor can ask their own agent to claim the highest
priority ready task matching their skills and interests. They continue steering
their agent and contribute the resulting work and evidence.

# Responsibilities

The coordinator should:

* query ready work from the Work Graph;
* match task requirements to an actor's declared and observed capabilities;
* respect owner policy, budgets, access boundaries, and concurrency limits;
* claim and renew the Work Graph's lease on behalf of the actor;
* assemble the smallest context package needed to begin;
* checkpoint progress before time or usage runs out;
* release work with a usable handoff; and
* record execution events for later analysis without filling the public DAG
  with raw telemetry.

# Personal mode

The first coordinator can be small. It only needs to know which of my agents are
available, what classes of work I trust each one to perform, and which ready
task is most useful. It should make resuming after a usage limit routine rather
than forcing the next session to rediscover the plan.

# Contributor mode

A public contributor should be able to request work using their own agent. The
request can include interests, skills, available tools, budget, and desired
complexity. The coordinator returns a task and a bounded context package. The
contributor's agent claims the lease, works in the contributor's environment,
and returns evidence or a proposed change through normal project controls.

This creates a more useful entry point than a stale "good first issue". The
work can match what the contributor wants to learn or support, while the graph
explains why it matters and which evidence will prove completion.

# First useful release

1. Wait until the Work Graph contains real ready tasks and reliable leases.
2. Define a portable description of task requirements and actor capabilities.
3. Add advisory matching for my own agents, with the final choice visible before
   claiming a lease.
4. Add checkpoints and clean handoff when an execution session ends.
5. Learn from completed, abandoned, and reassigned work instead of trusting
   model labels alone.
6. Add automatic matching within narrow owner-defined limits.
7. Test contributor-directed matching only after the personal and team modes
   work.

# Success

The coordinator succeeds when ready work spends less time waiting, actors receive
tasks they can finish, and an expired session does not destroy its progress. It
fails if it keeps cheap agents busy on low-value work or prizes subscription
capacity over project outcomes.

# Non-goals

* Owning project priority, dependencies, or completion state
* Receiving another person's subscription credentials
* Pretending a model name is a stable measure of capability
* Assigning every ready task when leaving capacity free is safer
* Removing the project owner's authority over what may run

## Initiatives

- [Semi-autonomous Software Development](https://robbiepalmer.me/initiatives/semi-autonomous-software-development.md): A connected system that keeps useful work moving, asks for human judgment when it matters, and learns how to need less intervention next time

---

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