Eric Do Couto · The playbookUpdated Oct 1, 2026

Work · Operating model · Sep 2026–now

Two agents, one GTM team

Codex and Claude Code share one task board for GTM work: 384 tasks from Sep 4 to Oct 1, 2026. Each task names one agent to do it and a different reviewer to check it.

Two keys

The method
First key
Codex or Claude Code claims the task, does the work and publishes the result.
Second key
A different reviewer, another agent or me, checks it against the acceptance criterion before it ships.

Built withCodex, Claude Code, Grok, MCP, Obsidian

  1. 384tasks logged on the agent hub, Sep 4–Oct 1, 2026On the record
  2. 13 → 0material findings over four review rounds on one handoffOn the record

The problem

Two AI agents do hands-on GTM work for me at Visualping: Codex and Claude Code. They write SQL against the data warehouse, draft posts for the CMS, build n8n workflows and audit Google Ads, HubSpot and Customer.io. Each agent works in its own terminal session, and the session’s context ends when the session does.

A single agent on a task holds both keys. The model that drafts the work also decides when it is finished, so its blind spots pass straight into the result. Two agents add a second risk: neither can see what the other has open.

The system

The agent hub is a shared task board, inbox, wiki and artifact store for the Codex and Claude Code sessions on my Mac. It runs as an MCP server and went live on Sep 4, 2026. Grok joins occasionally. Its wiki keeps decisions with their sources, so a decision outlives the session that made it. Every shared task moves through five steps:

  1. An agent opens a task for work I authorized, with one assignee, a different reviewer and a checkable acceptance criterion.
  2. The assignee claims the task for 30 minutes and renews the claim while it works.
  3. The assignee publishes the result as an artifact, versioned with a checksum.
  4. The assignee moves the task to review and tells the reviewer what to check.
  5. The reviewer approves, which marks the task done, or returns it with requested changes.

The agent hub’s skip rules keep small work off the board: short questions, status lookups, formatting and small edits that a deterministic check already verifies. A second agent joins only with a concrete job, such as a parallel subtask or a check on a subtle correctness risk.

One coordination rule on the agent hub came from a collision between Codex and Claude Code. Before changing a shared production system, an agent reads the board for an open task on that system and messages its owner first. A correct change can still break another agent’s plan when that plan assumed the old setup.

The two keys

Two-key GTM is the method behind the work on this site: nothing ships on one key. On the agent hub, the first key is the assignee and the second is the assigned reviewer, whose recorded approval is the only way a task gets marked done.

A bounded Claude review is the agent hub’s built-in second key. Codex can start a fresh Claude Code session that sees only the assigned task and its artifacts and can’t run code or browse. It gets eight turns, its brief caps feedback at 180 words, and its verdict covers only the version it read.

Review spend on the agent hub is tracked and rationed. Every bounded review leaves a receipt with token use, turn count, a cost estimate, and the tools it was given next to the tools it used. Each result gets one focused review, and a second needs a material revision or an unresolved finding.

A person can also hold the second key, recording the same verdict on the agent hub’s dashboard. Production systems get a stricter rule. For Google Ads, HubSpot, Customer.io and n8n, one agent holds the only write path, any change needs my approval given within the previous 30 minutes, and every change goes into a journal.

What it produced

The agent hub logged 384 tasks between Sep 4 and Oct 1, 2026. By Oct 1 it also held 561 versioned artifacts, 355 inbox messages (most of them between agents), 92 wiki note versions and 49 recorded reviews. Those tasks touch Google Ads, HubSpot, Customer.io, n8n, the data warehouse and the CMS.

In practice, Codex does most of the building on the agent hub and Claude Code most of the checking. Codex is the assignee on most tasks, and Claude Code recorded most of the 49 reviews.

The clearest second-key result came before the agent hub opened. On Aug 25, 2026, Claude and a GPT model running in Codex reviewed a colleague’s engineering handoff for the Visualping blog’s author pages: two code patches and a migration script. The findings the reviewers rated material fell from 13 to 6 to 1 to 0 over four rounds. In round one, the GPT model found eight that Claude had missed. Round three caught a rollback that sorted backups by filename and would have restored the wrong baseline. The migration ran two days later.

The agent hub opened on Sep 4, 2026, too recently to credit it with any change in signups.

What I’d change

Review becomes the constraint once agents produce at speed. Every task waits for its second key, so the reviewer’s time sets the pace of the whole agent hub.

My next step is to route the second key by risk. Reversible internal work, such as analysis, drafts and audits, will get an agent reviewer: the other model. I will review only what ships, spends money, sends a message or changes a production system.

A weekly review slot for the agent hub’s human queue comes next, so work that needs a person gets a fixed time to clear.