Expose durable agent resume APIs

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

@@ -36,13 +36,46 @@ your stack — the harness *is* the stack.
 | Discovery & RPC | Registry + client; agents and services find and call each other | Shipped |
 | Interop | MCP (tools), A2A (agents), x402 (paid tools) | Shipped |
 | Resilience | Per-call timeout with context propagation; opt-in retry/backoff (`ModelRetry`) across the loop | Shipped |
-| Durable runs | Checkpoint and resume an agent run (flows already do) | In progress |
+| Durable runs | Checkpoint and resume an agent run with the same checkpoint backend flows use | Shipped |
 | Observability | `RunInfo` → OpenTelemetry spans for runs, model calls, tools, delegation, and failures; persisted run history | Shipped |
 | Streaming | `ai.Stream` through chat, agent, and A2A | In progress |
 
 The "in progress" rows are exactly the roadmap's [Now and Next](/docs/roadmap.html),
 and the work is happening in the open.
 
+## Durable agent runs
+
+Agents can persist their execution history to the same `Checkpoint` backend as
+flows. A checkpointed `Ask` records the run id, original prompt, model result,
+and completed tool calls. If the process restarts after a tool succeeds but
+before the model finishes, `AgentResume` continues the same run and returns the
+recorded tool result instead of re-running the side effect. If a run already
+completed, resume returns the persisted response without calling the model.
+
+```go
+agent := micro.NewAgent("conductor",
+    micro.AgentProvider("anthropic"),
+    micro.AgentWithCheckpoint(checkpoint),
+)
+
+resp, err := agent.Ask(ctx, "charge order 42 and send a receipt")
+if err != nil {
+    // On startup, or after a transient failure, discover unfinished work:
+    pending, _ := micro.AgentPending(ctx, agent)
+    for _, run := range pending {
+        _, _ = micro.AgentResume(ctx, agent, run.ID)
+    }
+}
+_ = resp
+```
+
+For human-in-the-loop runs that pause through the built-in `request_input` tool,
+resume with the operator's response:
+
+```go
+_, err := micro.AgentResumeInput(ctx, agent, runID, "Deploy to us-east-1")
+```
+
 ## Observing agent runs
 
 Pass an OpenTelemetry tracer provider when you construct an agent to turn the

internal/website/docs/roadmap.md

@@ -36,14 +36,13 @@ The priority is that what exists works everywhere, under real conditions.
 
 ## Next — agentic depth
 
-- **Durable agent loop.** Flows resume; the agent's own loop does not yet. Reuse `Checkpoint` so a long-running agent survives a restart and continues.
 - **Streaming.** Broaden provider-backed `ai.Stream` coverage and keep chat plus A2A `message/stream` working end to end for real chat and long-task UX.
 - **Agent observability.** Wire the new `RunInfo` into OpenTelemetry spans so a run — steps, tool calls, delegation — is traceable. This is also what anyone running it in production will need.
 
 ## Later
 
 - **Memory management** — summarization and retrieval (RAG) beyond a fixed buffer.
-- **Human-in-the-loop** — pause and resume mid-run (`input-required`), beyond the binary `ApproveTool` gate.
+- **Human-in-the-loop** — broaden pause/resume UX around `input-required` runs and approvals.
 - **A2A** — richer live-stream reconnection (`tasks/resubscribe`) and `input-required` handoffs.
 
 ## Developer experience (ongoing)

micro.go

@@ -23,12 +23,18 @@ type Service = service.Service
 // Agent is the interface for an AI agent that manages services.
 type Agent = agent.Agent
 
+// AgentResponse is what an agent returns from Ask or a resumed run.
+type AgentResponse = agent.Response
+
 // AgentOption configures an Agent.
 type AgentOption = agent.Option
 
 // Flow is an event-driven LLM orchestration unit.
 type Flow = flow.Flow
 
+// FlowRun is a checkpointed flow or agent run record.
+type FlowRun = flow.Run
+
 // FlowOption configures a Flow.
 type FlowOption = flow.Option
 
@@ -164,6 +170,26 @@ func AgentWrapTool(w ...ai.ToolWrapper) AgentOption {
 // tool calls, delegation, and failures.
 func AgentTraceProvider(tp trace.TracerProvider) AgentOption { return agent.TraceProvider(tp) }
 
+// AgentWithCheckpoint sets the durability backend for agent Ask runs.
+// It uses the same Checkpoint interface as flows so services, agents,
+// and workflows can share one execution history backend.
+func AgentWithCheckpoint(c Checkpoint) AgentOption { return agent.WithCheckpoint(c) }
+
+// AgentPending returns checkpointed agent runs that have not completed.
+// Use it at process startup to discover agent work that should be resumed.
+func AgentPending(ctx context.Context, a Agent) ([]FlowRun, error) { return agent.Pending(ctx, a) }
+
+// AgentResume resumes a checkpointed agent run by id. Completed runs return
+// the persisted response without calling the model or replaying tool calls.
+func AgentResume(ctx context.Context, a Agent, runID string) (*AgentResponse, error) {
+	return agent.Resume(ctx, a, runID)
+}
+
+// AgentResumeInput resumes a checkpointed agent run waiting for human input.
+func AgentResumeInput(ctx context.Context, a Agent, runID, input string) (*AgentResponse, error) {
+	return agent.ResumeInput(ctx, a, runID, input)
+}
+
 // NewFlow creates an event-driven LLM orchestration unit.
 //
 //	f := micro.NewFlow("onboard-user",