Skip to main content
NOV 2026 Latest guide updates. Latest: AI Design Patterns Local vs cloud AI Grill me Changelog →
C14 Advanced Design

Agent Harness Map

Choose a runtime, repository contract, or orchestrator

PDF editions Β· Guide v3.44.1 Β· Updated

The four layers

LayerResponsibilityDoes not replace
ModelProposes text and tool callsPermission and recovery
Runtime harnessRuns the loop and toolsRepository contract
Repository harnessInstructions, state, verificationRuntime loop
OrchestratorCoordinates runtimes and sessionsEach runtime’s decisions

An orchestrator can launch several runtimes without owning their loops. A repository can improve its contract without changing runtime.

Loop and graph are not extra layers. Loop names the repeated feedback and stopping rule. Graph names the executable nodes, branches, joins, and checkpoints inside a runtime or an orchestrator.

A harness optimizer sits outside this stack. It changes and evaluates candidate harnesses for later runs.

Choose in the right order

  1. Define the task and acceptable autonomy.
  2. Select a runtime that owns the required loop.
  3. If the flow must be versioned, specify nodes, branches, and checkpoints inside the chosen runtime.
  4. Write the repository contract: instructions, pass criteria, verification.
  5. Add an orchestrator only when coordination, isolation, or recovery across tasks requires it.

Five sources, five uses

SourceUsed toLimit
Strict comparisonIdentify runtimesNot a universal ranking
Orchestrator tableCompare coordinationNot a runtime
Best of Agent HarnessesExplore 160 projects, 12 categories2026-08-23 snapshot
agent-harnesses.jsonRead provenance and evidenceunknown is not none
Optimizer research tableCompare search protocolsMostly research systems

Statuses for observed mechanisms: confirmed = visible mechanism, claimed = source assertion, unknown = not established yet, not_applicable = category mismatch.

Short trial, traceable verdict

Compare two or three model-harness pairs on real tasks in separate worktrees. Repeat critical tasks. Measure human verdict, interventions, plan drift, wall time, accepted-task cost, recovery, and setup friction.

Links: Agent Harness Map Β· Loop & Graph Engineering Β· Agent Harness Engineering Β· Agent Tools Β· Glossary

Enter your email to read the full card and get the complete PDF bundle.

All content is free and open-source. We just ask for your email.

PDF: