HAMMAD YOUSUF

FREE RESOURCE · FROM THE "HARNESS" REEL

Harness engineering: the six layers that make an AI agent reliable, with a checklist for each

By · 1 November 2026

The model suggests actions; the harness decides how, and whether, they run.

In four lines

  • The model suggests actions; the harness decides how, and whether, they run.
  • Check permissions in the system before any action runs, and give tools the least access they need.
  • Verify the real result (tests, records, payments) instead of trusting the model's reply.
  • Trace every tool call, error and retry, and feed what you learn back in.
The HARNESS reel: the harness diagram, one layer at a time. Watch on Instagram

The layers around the model

Agent failing? Don't blame the model yet. Walk the layers in the order a request passes through them. The first six are the ones the audit prompt names; Trusted Result and Rules & Limits are the two frames around them on the diagram.

Layer 01

Context Builder

give the model the right docs, data, history and task state before it starts.

Without it

The model guesses. It cannot see your calendar, your database or the last step unless the harness hands them over, so a missing fact becomes a confident wrong answer.

Checklist

  • The model sees the task state, not just the latest message
  • The docs and data it needs are in the context, not assumed
  • Earlier tool results are fed back in, in order
  • Nothing sensitive is in the context that the task does not need

Layer 02

Model

it reads the goal and suggests the next step; it can be wrong.

Without it

Nothing runs. But treating the model as the whole system is the mistake: it suggests, it does not decide, and it will mark work done that it has not tested.

Checklist

  • The model's output is treated as a suggestion, not a result
  • Every suggested action goes through the layers below before it runs
  • A model change is measured against fixed test cases, not vibes

Layer 03

Permission Check

every suggested action is allowed, blocked or sent to a human.

Without it

The model's text becomes an action with no gate in between. A wrong file path, a wrong recipient or a wrong amount runs exactly as written.

Checklist

  • Permissions live in the system, not in the prompt
  • Each action is one of: allowed, blocked, needs a human
  • Sends, payments, deletes and signatures always stop for sign-off
  • A blocked action is logged with the reason

Layer 04

Tools & Runtime

tools run with only the access the task needs.

Without it

A tool that can reach everything turns one bad call into a wide blast radius. Least access keeps a mistake inside the task that made it.

Checklist

  • Each tool has the least access the task needs (one folder, one account, read-only where possible)
  • Tools time out and fail loudly instead of hanging
  • A step cap stops a failing call from repeating

Layer 05

Verification

check the real result (tests, records, payments) before calling it done.

Without it

The agent reports success because the model said so. Anthropic's own harness work found agents tend to mark features complete without proper testing unless the harness forces an end-to-end test first.

Checklist

  • Done means a check passed, not a reply that said done
  • The check reads the real thing: the test run, the record, the payment status
  • One feature or task at a time, verified before the next starts
  • Progress is written to a file or history the next run can read

Layer 06

Trusted Result

only results that pass the checks leave the harness.

Without it

Unchecked output reaches users, customers or the next agent and the error travels further than the harness can see.

Checklist

  • Results carry what was checked and when
  • A failed check never leaves as a result; it leaves as a flag
  • The next system in line can tell a verified result from a draft

Layer 07

Observability

log tool calls, errors and retries, and feed the lessons back.

Without it

You cannot debug what you cannot see. A retry loop that costs money, or a tool that fails one time in ten, stays invisible until the bill or the complaint arrives.

Checklist

  • Every tool call, error and retry is traced with its inputs and outputs
  • Retries are counted and capped
  • A failure becomes a new test case or a new rule, not just a log line

Layer 08

Rules & Limits

the rules live in the system, not in the model.

Without it

A rule written only in the prompt can be forgotten, overridden or talked around. A rule enforced by the harness holds whatever the model writes.

Checklist

  • Guardrails are code, not instructions
  • Stopping conditions are set: steps, time, spend
  • Human checkpoints exist for the risky actions

Audit your agent's harness

Replace the bracketed part with a plain description of your agent and paste it into Claude or any assistant.

Audit your agent's harness

prompt
Here is how my AI agent works: [describe the agent, its tools and what it does]. Audit its harness against six layers: context builder, model, permission check, tools and runtime, verification, observability and feedback. For each layer: what I have today, the biggest risk if it fails, and one concrete fix I can ship this week. Finish with the three fixes that would most improve reliability, in order.

Common questions

What is harness engineering?

Designing the layer of code around an AI model that feeds it context, runs its tools, checks permissions, verifies results and logs everything. The model suggests actions; the harness decides how, and whether, they run.

Why not just use a better model?

A better model still only suggests the next step and can still be wrong. Reliability comes from the system around it: permissions checked before any action, results verified against the real thing, and traces that show what happened.

What does verification mean for an agent?

Checking the real result before calling the task done: run the tests, read the record, confirm the payment status. Anthropic's harness for long-running agents has agents test features end-to-end and mark them done only after testing, because they otherwise tend to mark work complete without proper testing.

Where does MCP fit in the harness?

In the Tools & Runtime layer. MCP servers expose tools, resources and prompts to AI applications, so the harness can give the model a standard set of tools while still deciding which calls are allowed and with what access.

How do I use the audit prompt?

Replace the bracketed part with a plain description of your agent, its tools and what it does, then paste it into Claude or any assistant. It returns what you have today, the biggest risk and one fix per layer, then the three fixes to ship first.

Sources checked

  • Anthropic's harness for long-running agents has agents test features end-to-end and mark them done only after testing, because agents otherwise tend to mark features complete without proper testing; it tracks progress in a file plus git history and works one feature at a time. Source ↗
  • Agents are LLMs using tools in a loop based on environmental feedback; they need guardrails, stopping conditions and human checkpoints. Source ↗
  • MCP servers expose tools, resources and prompts to AI applications. Source ↗

About the video: the reel has no presenter; it is an animated diagram. The small headshot in it is an AI-generated image of Hammad Yousuf.

Keep going