AdClaw v2: Coordinator Persona - your AI team now manages itself

I just shipped for adclaw.app something interesting that changes how AI agent teams work. Not another router. Not a prompt chain. A coordinator agent that synthesizes context across your entire team and makes decisions.

The problem nobody talks about:

You have @seo-expert running audits. @content-writer producing articles. @ads-manager launching campaigns. Each one is great in isolation. But nobody connects the dots.

SEO audit finds 5 pages with missing meta descriptions. Content writer has no idea. Ads manager keeps sending traffic to pages with broken H1 tags. This is the default state of every multi-agent system today. Agents work. Agents don't collaborate.

How the coordinator works

Runs on a cron schedule. Each cycle:

1. Reads AOM (Agent Object Memory) — what did each persona produce since last cycle?

2. LLM synthesizes cross-persona context with specificity enforcement — not "optimize further" (banned in prompt), but: "SEO found H1 issues on 3 pages. Blog post missing internal links to product pages."

3. Creates TaskStrategy with concrete delegations: → @seo-expert: list exact URLs of 5 pages missing meta descriptions → @content-writer: add 3-5 internal links to /pricing and /features

4. Delegates automatically. No human in the loop.

5. Next cycle — checks outcomes. If persona is stuck (same error twice), coordinator pivots. After 3 failed pivots — abandons with explanation.

Why this matters technically

Most "multi-agent" frameworks are glorified function routers. No memory, no cross-context reasoning, no feedback loops.

The coordinator has persistent memory via AOM, specificity enforcement at prompt level, outcome tracking across cycles, and failure recovery with pivot logic. Closed-loop coordination, not open-loop dispatch.

Real output:

Cycle 1: coordinator connects SEO audit (missing metas, duplicate H1s) with content output (SEO score 72/100) → delegates fixes to both personas.

Cycle 2: after results land, pushes content SEO from 72 to 85+ via meta, alt text, slug optimization. Zero human intervention.

The bigger picture

AdClaw is the infrastructure layer for autonomous AI teams. 118 skills. Persistent memory. Multi-persona orchestration. And now — self-managing coordination.

Open source. Apache 2.0.

github.com/Citedy/adclaw · pip install adclaw · adclaw.app