Harness Operations

Understand, compare, and benchmark the systems that turn models into agents.

Models get the attention. Harnesses determine how agent work actually runs: state, tools, permissions, delegation, recovery, and evidence. Harness Operations makes those differences visible.

Three illustrated ways people use AI: an engineer building software, an entrepreneur running a business, and a grandmother chatting with an AI assistant. Different goals. A more human future.

Current evidence snapshot

How much of the capability set is visible today?

15 operational capabilities, counted for each reviewed harness interface. This is capability coverage, not a performance benchmark.

Open full comparison →

AvailablePartialUnsupportedUnknownN/A

Anthropic Claude CodeClaude Code CLI · 2.1.282

12available

OpenCodeOpenCode CLI / headless server · 1.18.32

10available

CursorCursor CLI · hosted docs observed 2026-09-25

9available

OpenAI CodexCodex App Server · 0.157.0

9available

Harness Ops Bench is next.

The benchmark will measure repeated task completion, operational violations, recovery, cost, latency, and evidence quality—not just feature count.

Benchmark plan →

Use the reference

Understand a systemSee how Codex, Claude Code, Claude Projects, Browser Use, Computer Use, realtime agents, evaluation runtimes, and adjacent systems actually operate.

Compare capabilitiesInspect exact interfaces and versions across lifecycle, coordination, governance, isolation, resources, and evidence.

Follow Harness Ops BenchThe next step: reproducible trials measuring useful completion, operational reliability, recovery, cost, latency, and evidence quality.

Different shapes, different boundaries

A harness is not one architecture. These examples show why placement and orchestration matter.

Claude ProjectsOne coordinator, parallel Claude Code workers, and configurable cloud or machine execution.

Computer UseThe same computer-control capability can live in the client, a hosted runtime, or an outer Harness.

Realtime VoiceStreaming, interruption, transport, and live tool calls change the execution lifecycle.

Agent EvaluationAn evaluator can add sandboxing, approvals, recovery, and scoring that must not be credited to the Harness under test.

Go deeper when you need to

The Reference Model is a descriptive lens, not the prerequisite for using Harness Operations.

Operating arrangementsStandalone execution, embedded specialists, coordinator–worker threads, shared-state workers, handoffs, and independent services.

Standards and BoundariesWhere MCP, ACP, A2A, Code Mode, telemetry, and adjacent standards fit—and where they do not.

Reference ModelConcepts for execution, identity, evidence, governance, authority, and intervention.