# Agent Coordinator pitch

> Turn subsidized reasoning capacity into continuous progress on critical work

- HTML version: https://robbiepalmer.me/projects/agent-coordinator/deck
- Project: Agent Coordinator (https://robbiepalmer.me/projects/agent-coordinator.md)
- Slides: 9

## Slide 1

# Keep valuable agent work moving

Agent Coordinator puts available reasoning capacity onto the most valuable
automatable software-development work and carries it through a verified
outcome.

## Slide 2

## The scarce resource is human judgment

The first pilot launched general Work Graph tickets. It found plenty of work,
but much of it involved exploratory architecture, UX, or prose that still
needed a person to judge the result.

Cheap agent output becomes expensive when it only fills a human review queue.

## Slide 3

## Coordinate the work agents can finish

```mermaid
flowchart LR
Signal[Signal or ready work] --> Select[Select a verifiable outcome]
Select --> Execute[Launch and direct agents]
Execute --> Verify[Checks, delivery and QA]
Verify -->|proved| Done[Record the outcome]
Verify -->|recoverable| Execute
Verify -->|needs judgment| Attention[Ask for precise human input]
```

Selection, continuity, recovery, and evidence are one coordination problem.

## Slide 4

## Keep planning, coordination, and execution separate

### Work Graph

Desired outcomes, priority, dependencies, leases, and durable evidence.

### Agent Coordinator

Admission policy, capacity allocation, continuity, authority, and outcome evaluation.

### Agent runtime

Providers, sessions, threads, worktrees, limits, and interruption recovery.

The plan stays durable while execution machinery and model capacity change.

## Slide 5

## Use the capacity already available

* Match task requirements to capable agents and models
* Notice newly available work and safe parallelism
* Track subscription limits and reset times
* Resume useful work after limits, restarts, and interruptions
* Avoid launching work whose result still needs open-ended human review

Measure completed outcomes, not active threads or consumed tokens.

## Slide 6

## T3 Code replaces custom execution machinery

T3 Code already supplies signed webhook automations, controller and worker
threads, managed worktrees, provider sessions, pull-request watches, usage-limit
recovery, and continuation after restarts.

Build coordination policy and outcome loops instead of maintaining a second agent runtime.

## Slide 7

## Start with severe vulnerability remediation

GitHub provides the security signal. Renovate provides deterministic update
proposals. Agent Coordinator owns the judgment and delivery work that remains.

```mermaid
flowchart LR
Alert[Severe alert] --> Inspect[Inspect alert and Renovate PRs]
Inspect --> Fix[Repair or create the fix]
Fix --> Merge[Pass checks and merge]
Merge --> Deploy[Deploy]
Deploy --> Validate[Validate production]
Validate -->|failure| Fix
```

Security remediation is the first product vertical. The project covers the wider automatable SDLC.

## Slide 8

## Give the agent enough authority to finish

The worker may repair or replace a Renovate pull request, merge after required
checks, follow the established deployment path, validate production, and repair
or revert a failed rollout.

It asks for help when authority is missing or the safe result remains ambiguous.

A pull request is intermediate. The product is a verified outcome.

## Slide 9

## Prove one vertical, then expand across the SDLC

After vulnerability remediation, apply the same coordination to failing
delivery workflows, production regressions, dependency migrations, stale
high-value pull requests, and codebase analysis followed by bounded refactoring.

Expand when the next task class has an objective completion test.

[Read the full project account](/projects/agent-coordinator?tab=overview)

---

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