CodeHerder vs Claude Code Agent Teams
An experimental Anthropic feature. A Claude Code team lead starts teammates, each with its own context window. A shared task list and a mailbox link them.
- 2 November 2026
- Last checked
- 20
- Capabilities scored, side by side
Verified against the vendor's own published materials.
What Claude Code Agent Teams ships
- A team lead that starts teammates, each with its own context window
- A shared task list with automatic dependency tracking
- A direct mailbox between teammates. It works alongside subagents
- Hooks that can block a task from starting or completing. They run on your machine
Where Claude Code Agent Teams is strong
- It suits problems that one session cannot hold: codebase-wide review, multi-module features, and debugging with competing theories
How CodeHerder differs
- It works at a different level. One team solves one problem and then ends. CodeHerder keeps a durable backlog for many tenants over the life of an organization
- Anthropic says agent teams use significantly more tokens, and about 7x more in plan mode. CodeHerder tracks and limits that cost per task and per stage
Agent Teams solves one problem with a short-lived team. CodeHerder runs a backlog that lasts.
CodeHerder vs Claude Code Agent Teams, capability by capability
Every row traces to the same source as the full matrix, where the full legend lives. In short: "not published" means no public evidence was found, not that the feature is absent. Verified as of 2026-11-02.
| Capability | CodeHerder | Claude Code Agent Teams |
|---|---|---|
| Concurrent agents | Published | Published |
| Multi-vendor harnesses | Published | Not published (as of 2026-11-02) [1] |
| Customer-owned machines | Published | Partial [4] |
| Multi-machine routing | Published | Not published (as of 2026-11-02) |
| Automatic work assignment | Published | Published |
| Capability-based routing | Published | Partial (human-picked tier) |
| Atomic claiming | Published | Partial |
| Dependencies | Published | Published |
| Custom workflows | Published | Not published (as of 2026-11-02) |
| Server-enforced gates | Published | Partial (local hooks) |
| Agent handoffs | Published | Published (mailbox) |
| Persistent shared memory | Published | Not published (as of 2026-11-02) |
| Live terminal / view | Published | Partial (optional tmux) |
| Human takeover / steer | Published | Published |
| Cost per task | Published (not Cursor sessions) | Partial (token usage) |
| Cost per stage/model | Published | Not published (as of 2026-11-02) |
| Hard $ cap | Published (checked periodically) | Not published (as of 2026-11-02) |
| Schedules | Published | Not published (as of 2026-11-02) |
| Webhooks / API / CLI | Published | Partial (CLI only) |
| Self-host full platform | Published (Enterprise) | Not published (as of 2026-11-02) |
- Factory (Droids), Devin, and Claude Code Agent Teams are each vertically integrated — the product's whole thesis is a single vendor's own agent(s), so "multi-vendor harness support" isn't just unpublished, it's outside the product's own stated shape. Still written as "not published" rather than a claim of proven absence, per this tree's own rule.
- Codex app, Claude Code Agent Teams, and Augment Intent each run locally on the operator's own single machine (a real customer-owned-compute story for that one machine) but do not publish a *fleet* model spanning multiple customer-owned machines the way CodeHerder, Orchestratia, Gas City, Warp Oz, Coder Agents, Stoneforge, and Factory do.
Sources
Verified against Claude Code Agent Teams's own site as of 2026-11-02.
- Anthropic's Agent Teams documentation
- Anthropic's Claude Code cost documentation
Try it on one repository for two weeks.
File a task. An agent builds it. A person reads every request. You see the cost of each task.