ContextQMD
Back to Go Micro

The Agent Harness

internal/website/docs/guides/agent-harness.md

6.3.03.2 KB
Original Source

The Agent Harness

The first wave of agent frameworks solved one problem: put a model in a loop with some tools. The harder problem is operating that loop — and that's what a harness is.

A harness is the runtime around an agent:

  • the tools it can call,
  • the memory it keeps,
  • the guardrails that bound it,
  • the workflows that trigger and structure it,
  • the state that survives a restart,
  • the observability to see what it did,
  • the services it depends on,
  • and the protocols other agents use to reach it.

Go Micro's bet is that this runtime is the one you already deploy. An agent is a service with a model inside; the harness is the distributed-systems machinery services already have. So you don't bolt a separate orchestration product onto your stack — the harness is the stack.

The pieces, and what they map to

Harness concernIn Go MicroStatus
ToolsEvery service endpoint is an MCP-callable tool from registry metadata — no extra codeShipped
MemoryStore-backed agent memory (AgentMemory), durable across restartsShipped
GuardrailsMaxSteps, LoopLimit, ApproveTool, tool wrappers — enforced at the call siteShipped
WorkflowsDurable flows; flow.Loop for run-until-doneShipped
Planning / delegationBuilt-in plan and delegate tools on every agentShipped
Discovery & RPCRegistry + client; agents and services find and call each otherShipped
InteropMCP (tools), A2A (agents), x402 (paid tools)Shipped
ResiliencePer-call timeout with context propagation; opt-in retry/backoff (ModelRetry) across the loopShipped
Durable runsCheckpoint and resume an agent run (flows already do)In progress
ObservabilityRunInfo → OpenTelemetry spans; run history on the CLIIn progress
Streamingai.Stream through chat, agent, and A2AIn progress

The "in progress" rows are exactly the roadmap's Now and Next, and the work is happening in the open.

Why services are the right substrate

An agent that does real work needs typed, discoverable, callable capabilities — which is what a service is. The harness is credible because of the service layer, not in spite of it:

  • Tools are services — endpoint metadata becomes the tool schema; an RPC executes the call.
  • Agents are services — they register, load-balance, expose Agent.Chat, and are reachable by other agents.
  • Workflows are code paths — use a flow when the path is known; hand off to an agent when it isn't.
  • Safety lives at execution — guardrails run on the one path every tool call takes.

When to reach for it

Use Go Micro when the agent has to operate a system, not just answer a prompt — when it needs real tools, state that survives, limits you can enforce, and a way to be seen and called. If you only need a model in a loop, you don't need a harness. When that loop has to touch production, you do.

See also