ANATOMY OF AN AI CODING AGENT

The Harness
Is the Product

30-DAY TRIAL · PROFESSIONAL GUIDANCE INCLUDED

SLIDE TIME

6SEC

4 to 12 seconds, in steps of 1

Source: “The Harness Is the Product”, Towards AI · Theme: ampcode.com

WHERE THE ENGINEERING GOES: 85% HARNESS, 15% MODEL

SAME MODEL, WILDLY DIFFERENT OUTCOME

The harness is more important than the LLM

TEAM A · SAME MODEL + FULL HARNESS

Fixes bugs across a 100,000-file repo

Runs the tests · opens a clean pull request

VS

TEAM B · SAME MODEL, NO HARNESS

Forgets what you said 3 messages ago

Invents files that do not exist

LAYER 1 OF 6 · 5% OF THE PRODUCT

User Interface

IDE · Terminal · Web UI

Only 5% of the engineering effort
Start in the terminal, continue from the web or your phone
Agents keep working after you close the laptop

5%

amp > fix the flaky test, run CI, open a PR

✓ 41 tests passed · PR #128 opened · rollout monitored

TERMINALIDEWEBPHONE

LAYER 2 OF 6 · 15% OF THE PRODUCT

Agent Orchestrator

Planning & control loop

Breaks the goal into steps and picks the next move
Spawns parallel agents, each in its own fresh orb
Wakes itself on schedules and events, no baby-sitting

15%

LOOP

PLAN

ACT

OBSERVE

ADJUST

LAYER 3 OF 6 · 15% OF THE PRODUCT

Context & Indexing

Code search · repo indexing

Indexes the whole repository, not one file at a time
AGENTS.md carries project knowledge into every thread
Only the relevant slice enters the context window

15%

THE REPO100,000 FILES

INDEX + SEARCHEVERY THREAD, FRESH

IN CONTEXTTHE ~40 FILES THAT MATTER

LAYER 4 OF 6 · 20% OF THE PRODUCT

Tool Execution

File ops · build & test tools

The agent’s hands: it edits, builds and runs, not just talks
Sandboxed shell inside the orb, shared with you via tmux
MCP servers and plugins plug in any external system

20%

FILES

BUILD

TESTS

GIT

SHELL

MCP

LAYER 5 OF 6 · 15% OF THE PRODUCT

LLM Core: just the engine

The reasoning engine

The model every vendor fights about is 15% of the product
Swap the model, keep the harness: agents survive model churn
The other 85% is what makes the same weights feel magic

15%

LLM

HARNESS AROUND IT: 85%

LAYER 6 OF 6 · 20% OF THE PRODUCT

Verification & Feedback

Testing & analysis

The agent runs the tests, not you
Failures feed straight back into the loop
Proof before it ships: logs, screenshots, before/after numbers

20%

EDIT

→

BUILD

→

TEST

→

PASS

CROSS-CUTTING LAYERS · THE REMAINING 10%

Memory, State & Guardrails

MEMORY & STATE

Short-term memory · project knowledge

AGENTS.md · threads that persist · schedules the agent sets itself

SAFETY & GUARDRAILS

Sandboxing · access controls

Scoped credentials · approvals · every change auditable

5 + 15 + 15 + 20 + 15 + 20 + 10 = 100% HARNESS

30-DAY TRIAL · PROFESSIONAL GUIDANCE INCLUDED ampcode.com

The HarnessIs the Product