Skip to Content
GuidesHarness Engineering (advanced)

Harness Engineering on Exemplar

Advanced guide. Read this after Getting started with Marshal and SDKs & CLI.

Harness engineering is designing production agents from agent primitives—discover work, assign agents, verify results, and persist state—escalating to a human only when judgment is required. Exemplar’s Marshal, Tools & MCP, and Relay surfaces are those primitives, already wired for governance and audit.

Agent primitiveJobExemplar surface
SkillsPersistent proceduresSkill management
Memory / stateDurable spine across runsMemory
PromptsVersioned instructions and contractsPrompt management
Connectors / MCPTickets, CI, cloud, observabilityTools & MCP
VerificationMaker never grades its own workEvals · checker agents via frameworks
Guardrails & gatesBudgets, denylists, human approvalGuardrails · HITL · Relay for IDE
GatewayModel routing, cache, rate limitsAI Gateway
SchedulingCadence / triggersYour orchestrator or scheduler calling Marshal SDK / CLI (and Tools & MCP)

Practical mapping

  1. Discover — A cron, queue, or IDE skill starts a run; recall memory for open work and ownership.
  2. Act — Call governed MCP tools (harness.tools() or IDE MCP); keep secrets vaulted.
  3. Verify — Run evals or a separate checker agent; do not let the implementer grade itself.
  4. Gate — Use HITL for irreversible steps; use Relay Soft ask for IDE side effects.
  5. Persist — Write outcomes back to memory so the next run does not re-propose rejected work.

Autonomy phases

PhaseBehavior
Report onlyAgents summarize; humans act
ProposeAgents draft; HITL / Soft ask required
Act under capsAgents execute within guardrails and tool allowlists
TightenNarrow ask surfaces as confidence grows—do not remove gates silently

This page is a pattern guide for composing agent primitives—not a required onboarding step. Prefer Connect + one SDK hello-world first.

Last updated on