What it is

Harness engineering means wrapping models like Claude or Codex into an agent system that runs unattended — the three core pieces are loop (autonomous cycling), tools (tool calls), and guardrails. The model itself isn't "smarter"; what the harness builds is a closed-loop working system around it: where it runs, what it can call, what counts as "done," and how to recover on error. Both OpenAI and Anthropic have published official engineering notes ("Harness engineering: leveraging Codex" / "Effective harnesses for long-running agents"); Walking Labs turned it into a free open-source course.

Claude Duet is Anthropic's research-preview proactive agent released October 21, 2026 — not a new base model. It runs on Claude Opus 5.5, the same model behind Claude Code. The difference is the "contract": Duet is a cloud agent with its own VM + browser, connected to 6000+ apps you authorize (Gmail, Slack, Shopify…), currently in waitlist for US Pro / Max subscribers. It keeps working in the background while you're offline / have the app closed — reading mail, drafting replies, scheduling, running routines — and pauses for your approval on four action types: send message / pay / sign / delete.

One line on how they relate: harness engineering is the methodology; Claude Duet is one sample of that thinking productized. Understand how a harness is built and you can judge whether a "proactive agent" like Duet is actually reliable.

Hands-on / tutorial (Walking Labs free course)

Walking Labs' Learn Harness Engineering is an MIT-licensed, multilingual, paywall-free open course (GitHub repo ~21k stars). Its core claim: same model, a weak harness makes you the "cleanup crew," a strong harness makes you the "reviewer." The course breaks a harness into five subsystems — instructions, state, verification, scope, session lifecycle — and lands them in your repo via four files:

your-repo/
├── AGENTS.md          # operating manual for the agent (also called CLAUDE.md)
├── feature_list.json  # scope: feature list + done / not-done
├── claude-progress.md # session log (rename for Codex)
└── init.sh            # install + verify before work begins

Minimum three steps:

1. Drop the four files in your project root — the next agent session reads them first instead of guessing repo rules from scratch. 2. Run the audit script (zero dependencies, no Node needed):

bash <(curl -fsSL https://raw.githubusercontent.com/walkinglabs/learn-harness-engineering/main/tools/audit-harness.sh) /path/to/your/repo

The script checks all key harness items and lists "weak spots" like missing progress logs or missing verification hooks. 3. Project 01 A/B: run the same task twice — first with raw prompts, then with the harness files. Compare how many times the agent declared "done" before tests passed. Later lessons and projects layer harness onto an Electron knowledge-base app; the capstone ships with a benchmark and ablation (remove subsystems one by one to see what breaks first).

Comparison: Claude Duet vs Claude Code (and harness vs raw prompts)

How it differs from Claude Code — the point: same model, different boundary.

DimensionClaude CodeClaude Duet
ModelClaude Opus 5.5Claude Opus 5.5 (same)
Runs onyour machine / CI, next to the repoAnthropic cloud VM + browser
Work scopeinside the repo: read/write files, run commands, run testscross-app knowledge work: mail / docs / invoices / calendar
"Done" meansobjective: tests pass, CI passes, PR mergeablesocial: you judge if it's right
Approval guardraildeveloper supervises hands-onsend / pay / sign / delete need approval; securities trading and inter-account transfers rejected even if approved
For whomindividuals writing codesmall teams wanting to offload ops / scheduling / cross-tool work to an agent

One line: Claude Code is the engine; Claude Duet is the engine put into a car that "runs your business 7×24 for you." Duet doesn't replace Claude Code — code changes still belong in the repo, through tests and review, where Claude Code fits; Duet shines at the "still moving while you're away" kind of work.

> Don't confuse the same-named product: searching "Duet" surfaces a third-party SaaS duet.so ("cloud AI coworker") — that's a separate company, not Anthropic's Claude Duet. Keep the distinction when citing.

Harness vs raw prompts: prompts teach the model "how to think"; the harness decides "when it can act, where, how to stop, how to verify." The former is thrown away per chat; the latter persists context, scope, and verification to files — exactly the engineering premise that lets a long-running agent like Duet "keep going offline," and what Walking Labs' course actually teaches.

Sources

  • Anthropic official: Introducing Claude Duet (2026-10-21 research preview), Effective harnesses for long-running agents, Harness design for long-running application development
  • OpenAI: Harness engineering — leveraging Codex in an agent-first world
  • Walking Labs: Learn Harness Engineering (github.com/walkinglabs/learn-harness-engineering, MIT, ~21k stars)
  • Neutral comparison: OrcaRouter "Claude Duet vs Claude Code: one model, two definitions of 'done'"; media coverage of Duet's safety model (send / pay / sign / delete approval, virtual-card payment)

← Back to all guides