Orient
Load the goal, constraints, tools, and relevant context.
A field guide · v1.0
Give an LLM
a way to act.
An agent harness is the runtime around a language model—the tools, memory, feedback loops, and guardrails that turn a conversation into a capable system.
MODEL
reason
→ act
→ act
toolsworld access
memorycontext over time
loopobserve + adapt
THE HARNESS
LLM + tools + state + feedback✳≠ just a chatbot✳
LLM + tools + state + feedback✳≠ just a chatbot✳
01 / The idea
The model is the mind.
The harness is the agency.
A raw LLM can predict the next token. A harness gives it a job, a place to work, and a way to learn what happened after it acted.
contexttoolsstatepoliciesevaluation
The agent loop01 → 04
↗
Decide
Reason about the next best action, not the final answer.
Act
Call a tool in the real world: code, browser, API, or shell.
Reflect
Read the result, update state, and keep going—or stop.
02 / MCP
Model Context Protocol
is the tool contract.
MCP standardizes how a harness discovers tools, requests actions, and receives structured results. It makes tools portable across models and clients, so you can swap infrastructure without rewriting the agent loop.
Read the full MCP guide →01
Discovery
The harness can enumerate available capabilities and their schemas before planning.
02
Invocation
Tool calls are explicit and validated, reducing brittle prompt-level parsing.
03
Interoperability
One protocol supports many servers, enabling consistent cross-environment tooling.
03 / Skills
Skills let a harness
delegate with intent.
Skills package domain playbooks, tools, and policies into reusable capabilities. They help the harness route tasks to the right workflow instead of forcing one generic loop to handle every job.
Read the full Skills guide →01
Task fit
Each skill narrows scope and context, increasing precision for specific work types.
02
Safety
Skills can enforce scoped permissions and checks before high-impact actions.
03
Throughput
Specialization reduces reasoning overhead and speeds multi-step task completion.
04 / Catalog
Real systems.
Different loops.
A practical field guide to agent products, open-source runtimes, and model-specific harnesses.
CODING01
OpenAI
Codex
Cloud and local coding agents that inspect repositories, edit files, run checks, and hand back a change.
CODING02
Claude
Code
Anthropic's terminal-first coding agent with repository context, shell access, edits, and a reviewable loop.
CODING03
GitHub
Copilot
Agent mode and coding agent experiences that turn issues and prompts into changes inside GitHub workflows.
CODING04
Open
Code
An open-source terminal coding agent with model choice, tool calls, file edits, and a configurable workflow.
CODING05
Pi
coding agent
A minimal, extensible coding harness designed to keep the agent loop visible and hackable.
CODING06
OpenHands
An open-source software engineer agent that works through a runtime with a browser, shell, files, and feedback.
RESEARCH07
Hermes
A personal agent-style harness for connecting a model to tools, memory, and everyday tasks.
RESEARCH08
OpenClaw
An open agent harness pattern for giving a model a persistent workspace, tools, and a long-running task loop.
RESEARCH09
DeepSeek
Harness
A model-focused harness concept for putting DeepSeek reasoning behind tool use, planning, and verification.
RESEARCH10
Browser
Use
A browser-control runtime that lets an agent navigate sites, fill forms, and report what happened.
RESEARCH11
Perplexity
Research
A search-and-citation workflow that decomposes questions, browses sources, and synthesizes an answer.
RESEARCH12
LangGraph
A graph runtime for explicit agent state, branching workflows, tool calls, checkpoints, and human review.
AUTOMATION13
Goose
An open-source developer agent that connects models to local tools, APIs, and repeatable engineering workflows.
AUTOMATION14
AutoGen
A framework for composing agents that message one another, call tools, and coordinate multi-step work.
AUTOMATION15
CrewAI
A role-based multi-agent runtime for delegating research, planning, and execution between specialized agents.
AUTOMATION16
Cline
An IDE agent that can inspect a project, edit files, run commands, and ask for approval before actions.
AUTOMATION17
SWE-agent
A research harness for software-engineering tasks that turns issue context into tested repository changes.
AUTOMATION18
Operator
A computer-use style harness for completing everyday browser tasks with observation, action, and recovery.
05 / Blueprint
Build the
smallest loop.
Start with one job and one tool. Add complexity only when the feedback says you need it.
See the principles ↓job: ship a safe change
model: reasoner
tools:
- read_files
- run_tests
- write_patch
loop:
observe → decide → act → verify
stop_when: tests_pass && diff_is_smallThe good harness checklist
01
Bounded
Clear permissions, budgets, and stop conditions.
02
Observable
Every action leaves a trail you can inspect.
03
Correctable
Humans can pause, redirect, and undo the work.
04
Evaluated
Success is measured against the actual job.