agentd

A 3 MiB static binary that runs one supervised AI agent — MCP-native, no local code execution.

graduatedAgent runtimesApache-2.0
3.0 MiB
static binary
<1 ms
cold start
3
direct dependencies
579
tests + 38 conformance

agentd runs one agent as a ReAct loop: hand it an instruction and one LLM endpoint and it thinks, calls a tool, observes, and repeats until a terminal status. Every tool comes from a remote MCP server over HTTPS — agentd ships none of its own, and never executes local code. A model-free supervisor owns lifecycle, limits, and a killable process tree, so a runaway model is contained by construction.

Five run modes (once, loop, reactive, schedule, workflow) make it a cloud-native unit of work: drop it into a Kubernetes Job, a CronJob, or a long-lived reactive Deployment that idles at near-zero CPU and wakes on MCP resource changes. Agents compose like Unix processes — an agentd serves its own MCP, delegates over A2A, and nests subagents.

It is the reference implementation of the open Agent Control Contract (ACC v1), the same contract the agentctl control plane manages fleets through — by contract, never by code dependency.

Related projects