Build an application with two different responsibilities. A primary Loop owns the user-facing report. A researcher Loop loads citation instructions, searches a bounded corpus, and returns a cited evidence note. The application writes the primary Loop’s final Markdown into its workspace.

What you will build

The example proves delegation with a real Harness child Loop and SubmitToLoop, not a helper function pretending to be another agent. Its local article corpus makes retrieval and citation tests deterministic. The same tool boundary can later adapt an MCP search server.

Architecture

%%{init: {"theme":"dark"}}%%
flowchart LR
  U[User question] --> P[Primary Loop]
  P --> D[Child researcher Loop]
  D --> S[research-citations skill]
  D --> C[Local article corpus]
  C --> D
  D -->|cited evidence note| P
  P --> R[research-report.md]

The Rig applies a delegation depth and quota. Each Loop has its own model, instructions, tool surface, and skill allow-set. The researcher cannot acquire the primary Loop’s authority merely because it is a child.

Project structure

research-assistant/
├── corpus/
│   └── skills.txt
├── skills/
│   └── research-citations/
│       └── SKILL.md
├── go.mod
├── go.sum
├── main.go
└── main_test.go

The generated artifacts/research-report.md is runtime output and is not part of the source tree.

Build the researcher

researcher, err := loop.Define(
	loop.WithName("researcher"),
	loop.WithInference(researchClient, researchModel),
	loop.WithSystem("Load research-citations, search the corpus, and return a cited evidence note."),
	loop.WithTools(skillDefinition, searchDefinition),
	loop.WithAccessGate(accessGate),
	loop.WithPolicyRevision("research-worker-v1"),
)

corpus.search accepts a query and returns an array of {title, source, text} records. The embedded skill requires the researcher to search before drafting, keep claims within retrieved evidence, and cite every factual claim.

For project-specific instructions, also call skill.WithWorkspaceRoot(root) and advertise metadata from skill.DiscoverWorkspaceSkills(root). Keep workspace skill loading behind the workspace trust boundary.

Delegate the subtask

primary, err := loop.Define(
	loop.WithName("research-assistant"),
	loop.WithInference(primaryClient, primaryModel),
	loop.WithDelegates("researcher"),
)

runtime, err := rig.Define(
	rig.WithLoops(primary, researcher),
	rig.WithPrimers("research-assistant"),
	rig.WithSessionStore(sessions),
	rig.WithDelegationLimits(rig.DelegationLimits{Depth: 2, Quota: 2}),
)

childID, err := live.NewLoop(
	loop.Provenance{LoopID: live.ActiveLoop().ID()},
	researcher,
)
_, err = live.SubmitToLoop(ctx, childID, []content.Block{
	&content.TextBlock{Text: question},
})

The full example waits for the child Loop’s TurnDone, then submits its evidence note to the primary Loop. See Delegation for managed delivery, cancellation, restore, and authority limits.

Save the artifact

if err := os.MkdirAll(workspace, 0o755); err != nil {
	return err
}
if err := os.WriteFile(
	filepath.Join(workspace, "research-report.md"),
	[]byte(report),
	0o600,
); err != nil {
	return err
}

The test provides a unique temporary workspace and compares the complete report. A product can instead bind Harness Workspaces, checkpoint them, and restore both Session and artifact state. Add Compaction when long research threads approach their context limit.

Run it

From the standalone example directory:

GOWORK=off go test -race ./...
GOWORK=off go run .

The second command creates artifacts/research-report.md. Remove that directory when you no longer need the generated report.

Expected interaction

Researcher: Skills are loaded on demand and authorized per agent. [1]
Assistant: wrote a cited brief from 1 source.
Saved: research-report.md

The report contains the claim, citation [1], and a Sources section naming skill/skill_loader.go. The deterministic test rejects any drift in the file.

Use a live model

Replace the two scripted clients independently. The researcher can use a fast hosted or local model while the primary Loop uses a stronger model for synthesis. To retrieve live sources, replace corpus.search with an adapter built on the MCP module, keep its returned source metadata, and require network permission through Gates.

The example intentionally keeps live search optional. CI should continue running the local corpus path so citations and workspace output remain reproducible.

Try next

Add an exact evaluator that rejects uncited factual sentences.
Interrupt the researcher after retrieval and resume from persisted Session state.
Attach the TUI or Web UI to the same Session event and command interfaces.
Add a second researcher with a different source catalog and compare evidence before synthesis.

Research Assistant