New Cohort Starts:

Donate
← Back to the skill map
Reliability & Operations/AI & reliability

Harness Engineering

7 micro-topics in Reliability & Operations, each with the evidence that proves you have it and the written reason behind every link. Before you start, it rests on 6 other domains. Nothing in the map rests on it yet — it is a leaf.

7 topics  ·  depth 11–16 of 16  ·  6 internal links  ·  7 in  ·  0 out

Before you start

Load-bearing links first. Each one says what in this domain rests on what outside it, and why.

Agents

2 links

load-bearing

Distinguish a harness from an execution layer rests on Run an agent loop

A harness wraps a loop that runs.

load-bearing

Set retry and budget ceilings rests on Bound loops with budgets

The ceiling moves the loop's budget into the harness.

LLM Evals

1 link

load-bearing

Add an independent checker rests on Write deterministic checks

A checker needs a definition of correct.

Observability

1 link

load-bearing

Encode a harness after observed failure rests on Trace LLM applications

You encode failures you observed.

AI Safety & Governance

1 link

load-bearing

Write a policy layer rests on Defend against prompt injection

Untrusted input is the reason the policy exists.

Delivery & CI/CD

1 link

supporting

Define rollback boundaries rests on Roll back safely

Reversibility is the same discipline applied to agent actions.

Shell Scripting

1 link

supporting

Distinguish a harness from an execution layer rests on Handle exit codes and errors

Exit codes are the simplest pass-fail signal a harness reads.

What you will be able to do

In prerequisite order. Each idea names the artifact that closes it, the market demand that put it on the map, and what it stands on.

Depth 11  ·  conceptual  ·  scaffolded  ·  M32

Distinguish a harness from an execution layer

Knows what shell access alone does not give you

Evidence
Explains why bash access is not a harness
Market anchor
AI systems reliability
Rests on
load-bearing

Run an agent loop · Agents

A harness wraps a loop that runs.

supporting

Handle exit codes and errors · Shell Scripting

Exit codes are the simplest pass-fail signal a harness reads.

Depth 12  ·  conceptual  ·  scaffolded  ·  M32

Add an independent checker

Separates the maker from the judge

Evidence
Blocks a known failure mode without blocking correct work
Market anchor
AI systems reliability
Rests on
load-bearing

Distinguish a harness from an execution layer

The checker is the component the definition demands.

load-bearing

Write deterministic checks · LLM Evals

A checker needs a definition of correct.

Depth 12  ·  conceptual  ·  scaffolded  ·  M32

Set retry and budget ceilings

Bounds cost and looping at the harness level

Evidence
A failing agent stops instead of retrying forever
Market anchor
AI systems reliability
Rests on
load-bearing

Bound loops with budgets · Agents

The ceiling moves the loop's budget into the harness.

load-bearing

Distinguish a harness from an execution layer

Bounding is a harness responsibility.

Depth 13  ·  meta  ·  guided  ·  M32

Separate connective from compensatory harness

Knows which parts survive a model upgrade

Evidence
Labels each harness component and deletes a stale one
Market anchor
AI architecture
Rests on
load-bearing

Add an independent checker

You classify components you have built.

Depth 14  ·  meta  ·  guided  ·  M32

Encode a harness after observed failure

Builds on the second occurrence, not the first

Evidence
Points to the failure log that justified a component
Market anchor
AI architecture
Rests on
load-bearing

Separate connective from compensatory harness

Knowing which kind you are building tells you when to build it.

load-bearing

Trace LLM applications · Observability

You encode failures you observed.

Depth 15  ·  representational  ·  guided  ·  M32

Write a policy layer

Allowlists actions and sandboxes execution

Evidence
A disallowed action is refused before it runs
Market anchor
AI safety · security
Rests on
load-bearing

Add an independent checker

Policy decides what the checker's verdict permits.

load-bearing

Defend against prompt injection · AI Safety & Governance

Untrusted input is the reason the policy exists.

Depth 16  ·  conceptual  ·  guided  ·  M32

Define rollback boundaries

Makes agent actions reversible

Evidence
An agent change is undone cleanly
Market anchor
AI systems reliability
Rests on
load-bearing

Write a policy layer

Reversibility is what makes a permissive policy survivable.

supporting

Roll back safely · Delivery & CI/CD

Reversibility is the same discipline applied to agent actions.

What rests on this domain

Everything downstream that names an idea here as a prerequisite, grouped by where it lives.

Nothing in the map rests on this domain yet — it is a leaf.

Retool. Retrain. Relaunch.

375 ideas. 17 weeks. No tuition, ever.

Vets Who Code is a veteran-run 501(c)(3). The accelerator is free, remote, and we don’t take a share of your first paycheck.

Free · Remote · 17 weeks · EIN 86-2122804