Agent registry
Every agent is a versioned object: prompt, model, tools, budget and guardrails. Fork it, diff it, roll it back.
Claude agent operations
Command a whole fleet of Claude agents from one screen.
AnthropicSub is the agent manager for Anthropic models. Spin up agents, give them tools, route work between them, and watch every token, run and decision in real time.
- 2.4M
- agent runs / week
- 38ms
- routing overhead
- 99.98%
- run completion
Claude OpusClaude SonnetClaude HaikuMCP toolsSub-agentsStreaming runsBudget guardsAudit trailsClaude OpusClaude SonnetClaude HaikuMCP toolsSub-agentsStreaming runsBudget guardsAudit trails
The product
Everything an agent fleet needs, minus the duct tape.
Orchestration graph
Chain agents into graphs with fan-out, retries and human checkpoints. No glue scripts to babysit.
Tool layer
Attach MCP servers, HTTP tools or your own functions once, then share them across the whole fleet.
Live run console
Watch reasoning, tool calls and token spend stream in as they happen. Stop a run mid-flight.
Guardrails
Hard spend caps, allowed-tool lists and approval gates on anything that writes, pays or sends.
Evals on every change
Ship a prompt edit only when it beats the current agent on your own test set.
Live orchestration
See what every agent is thinking, right now.
One console for the whole fleet: status, queue depth, spend and the full reasoning trace of any run. Pause, redirect or kill an agent without redeploying anything.
fleet / production
research-scoutOpus
invoice-triageSonnet
support-routerHaiku
contract-readerSonnet
12:04:01 research-scout → tool:web.search("agent orchestration")
12:04:03 research-scout ← 8 results (1.2k tokens)
12:04:05 spawn sub-agent: summarizer (Haiku)
12:04:09 guardrail: spend 0.42 / 5.00 USD — ok
12:04:12 handoff → support-router
12:04:14 ▍
The workflow
Four steps from idea to a governed agent in production.
- 01
Define the agent
Pick a Claude model, write the system prompt, attach tools and set a hard budget.
- 02
Wire the graph
Connect agents with handoffs, parallel branches and human approval gates.
- 03
Run and watch
Stream every step live. Replay any run from any point with changed inputs.
- 04
Prove it got better
Evals, diffs and spend reports on every version before it reaches production.
Pricing
Priced per human, not per agent.
Solo
$0
for one builder
- 3 agents
- Live run console
- Community MCP tools
- 7-day run history
Most teams
Fleet
$79
per seat / month
- Unlimited agents
- Orchestration graphs
- Spend caps & approval gates
- Evals on every version
- 90-day run history
Enterprise
Talk to us
your cloud or ours
- SSO & SCIM
- Private networking
- Audit export
- Dedicated support