Build Harness

Building Agent and Harness from 0 to 1

A hands-on path from a simple CLI assistant to a controllable, observable Agent Harness.

Start reading

Contents

Read in order to move from model fundamentals into an engineering view of agents and harnesses.

01 Agent Base Definition: Why It Is Not a Prompt Guide When many people first start building Agents, the most natural reaction is: if we make the system prompt longer, write the rules in more detail, wi...
02 Agent Composition Model: Model, Loop, Tools, State In the previous article, we first removed one misunderstanding: an Agent is not a longer prompt.
03 System Boundaries: The Difference Between ChatBot, Workflow, Agent, and Harness When people first build Agent systems, they often naturally read them as an upgrade path:
04 Harness Base Definition: The Control System Outside the Model Previously, we split Agent into several minimal parts:
05 Agent Evolution Path: Chat Agent -> Tool Agent -> Runtime Agent -> Managed Agent When people first look at Agent architecture diagrams, a natural confusion appears: why does a "model that can chat" eventually become a whole Harn...
06 Why Write an Agent by Hand: Understanding the Minimal Mechanisms Behind Framework Abstractions Across the previous five posts, we've been doing one thing: pulling Agents back from "magical model capabilities" to "explainable runtime systems."
07 LLM Provider Integration: Let the CLI Make Its First Model Call The previous chapters have all been about the boundary between the Agent and the Harness.
08 Minimal Agent Loop: From One-Off Answers to Multi-Step Action In the previous chapters we kept circling around a single question: an Agent isn't a single Prompt, nor is it a chattier model — it's a running sys...
09 M0 Core Kernel: Wire Real LLMs into the System, Don't Let Them Take Over The previous articles have laid out the mental model for Agent and Harness.
10 Intent / Execution Separation: The Model Proposes, the System Executes A lot of people, when they first wrote CLI Agent, thought of it as a direct call:
11 Plugin Host: Why Must Core Learn to Be Extended? In Article 10, we defined an important boundary:
12 Provider Runtime: why can a provider only return tool intent? In the previous group of articles, we defined a low-level discipline:
13 Tool Runtime: from tool intent to observation In Article 10 we drew a clear boundary:
14 Local Tool Bundle: files, search, terminal, and permission runtime At this point, many people are tempted to model an Agent's local capabilities as a very intuitive set of functions.
15 Context Policy: what should the model see in this round? The previous articles have already split apart the Agent action chain.
16 Session Replay: why is the event log the source of truth for long tasks? When many people add persistence to an Agent for the first time, they naturally save messages.
17 Capability Discovery: Skills, MCP, and dynamic tool exposure By Article 17, our small CLI Agent is no longer the original chat-only program.
18 Delegation Runtime: delegate work without losing control At this point, our small CLI Agent is no longer just a chat-shaped model wrapper.
19 Trace Analysis: locating Agent failures with fact logs The previous chapters gradually pushed a small CLI Agent into a more realistic position.
20 Memory Governance: from candidate ledger to governance store By part 20, our small CLI Agent can already do quite a lot.
21 Scoped Retrieval: from bounded retrieval to audit snapshots Many people design the first retrieval layer for an Agent in a very direct way.
22 Productized CLI: profile, extension, multi-provider At this point, our small CLI Agent is no longer the early demo that could only run once.
23 Hosted Harness: Sandbox, Cron, Durable Execution, and remote deployment Move a local CLI Agent into a hosted environment and understand how sandbox, cron, durable execution, workspace setup, secret boundary, artifact store, resume/retry, and notification form a Hosted Harness.
24 Agent Harness terminology map: Intent, Observation, Event, Artifact, Snapshot, Projection, and Trace Starting with Tool Runtime, the series enters a stage where terms can easily blur together.
25 Building Agent and Harness from 0 to 1 A hands-on engineering tutorial that starts with a simple CLI assistant and grows it into a controllable Agent Harness.