AI harnesses turn language models into agents that can do real work. This overview presents the most important coding agents and orchestrators – factually, with a focus on features, model access, permissions and enterprise use.
AI Harnesses
Agent Harness
AI harnesses: the software that lets AI agents work
The model thinks, the harness acts. Claude Code, Codex, Pi and other agent harnesses give language models tools, context, permissions and an environment. This overview shows what each harness does and how companies use them in a controlled way.
What is an AI harness?
An AI harness – also called an agent harness – is the software around a language model that makes it able to work: it drives the agent loop, provides tools and an execution environment, manages context and memory, enforces permissions and coordinates multiple agents where needed. The same model performs differently in different harnesses – which is why the harness deserves as much attention as the model.
Read the primer: AI harness explained →
In focus: Pi, Claude Code and Codex
Three coding harnesses that are particularly common in development teams
Coding agent
Pi
Earendil Inc.Pi is the open-source, minimalist coding agent harness from Earendil Inc. for the terminal – extensible via TypeScript, with access to over 15 model providers.
Coding agent
Claude Code
AnthropicClaude Code is Anthropic's agentic coding harness for terminal, IDE, desktop and web, with tools, context, permissions and subagents for developers.
Coding agent
Codex
OpenAICodex is OpenAI's agentic coding harness for the terminal, IDE and cloud. Features, sandbox modes, model access and enterprise use at a glance.
Orchestration: running several agents in parallel
herdr runs coding agents in real terminals, shows their state and lets them drive each other
The seven building blocks of a harness
Every harness detail page is described along these seven building blocks
- 1
Agent loop
Calls the model, executes tool requests and decides when the task is finished.
- 2
Tools and execution environment
Files, shell, web and MCP servers – executed in a controlled environment, often a sandbox.
- 3
Context management
Instruction files, compaction and just-in-time loading keep the limited context window usable.
- 4
State and memory
Sessions, checkpoints and notes allow work beyond a single context window.
- 5
Permissions
The harness, not the model, decides which action may run without approval.
- 6
Verification and feedback
Tests, review steps and human approvals turn 'done' into 'done correctly'.
- 7
Orchestration
Sub-agents and parallel sessions split complex tasks into parts.
All AI harnesses in detail
Click a harness for features, model access and enterprise use
Coding agent
Amp
Amp Frontier CorporationAmp is the agentic coding harness by Amp Frontier Corporation (formerly Sourcegraph) for terminal, web, and mobile apps. Models, …
Coding agent
Claude Code
AnthropicClaude Code is Anthropic's agentic coding harness for terminal, IDE, desktop and web, with tools, context, permissions and …
Coding agent
Cline
Cline Bot Inc.Cline is Cline Bot Inc.'s open-source AI harness for VS Code, JetBrains, terminal and SDK: a coding agent with MCP, checkpoints …
Coding agent
Codex
OpenAICodex is OpenAI's agentic coding harness for the terminal, IDE and cloud. Features, sandbox modes, model access and enterprise use …
Coding agent
Cursor Agent CLI
Anysphere (Cursor)Cursor Agent CLI is Anysphere's terminal coding agent behind the Cursor editor. Models, permissions, MCP, checkpoints and …
Coding agent
Devin
CognitionDevin is Cognition's AI software engineer, combining a cloud agent with a local Devin CLI. Sessions, Knowledge, Playbooks, MCP, …
Coding agent
Droid (Factory)
Factory (The San Francisco AI Factory Inc.)Droid is Factory's coding agent harness for the terminal, desktop, and cloud. Autonomy levels, BYOK models, MCP support, and …
Coding agent
Gemini CLI
GoogleGemini CLI is Google's open-source coding agent for the terminal. Features, models, MCP, sandbox and enterprise use of the agent …
Coding agent
GitHub Copilot
GitHub (Microsoft)GitHub Copilot is Microsoft's AI assistant for developers, with Copilot CLI, coding agent and agent mode. Models, permissions, MCP …
Coding agent
Google Antigravity
GoogleAntigravity CLI is Google's agent-first terminal interface for coding agents with selectable Gemini, Claude and GPT-OSS models, …
Coding agent
Grok CLI
xAI (SpaceXAI)Grok CLI, officially Grok Build, is the open-source coding agent from xAI (SpaceXAI) for the terminal. Features, models, sandbox, …
Orchestrator
herdr
Herdr, Inc.herdr is the open-source terminal runtime by Herdr, Inc. for AI coding agents such as Claude Code or Codex, with state detection …
Coding agent
Hermes Agent
Nous ResearchHermes Agent is Nous Research's open-source, self-improving AI agent harness with memory, skills, a messaging gateway and free …
Coding agent
Kilo Code
Kilo CodeKilo Code is Kilo Code Inc.'s open-source AI harness for VS Code, JetBrains and the terminal: a coding agent with over 500 models …
Coding agent
Kimi Code
Moonshot AIKimi Code is Moonshot AI's open-source coding agent for terminal, web UI and IDE. Features, models, custom endpoints and …
Coding agent
Kiro
Amazon Web Services (AWS)Kiro is AWS' agentic coding harness for terminal, IDE and web, with spec-driven development, steering files, hooks, MCP and an EU …
Coding agent
Letta Code
LettaLetta Code is Letta's memory-first coding agent harness, with persistent, learning agents, memory blocks and sub-agents for …
Coding agent
Maki
tontinton (open-source project)Maki is an open-source (MIT) coding agent for the terminal, optimised for low context usage, with subagents, MCP support and broad …
Coding agent
Mastra Code
MastraMastra Code is Mastra's open-source terminal coding agent harness with Observational Memory, Build/Plan/Fast modes and broad model …
Coding agent
Muse Code
Meta PlatformsMuse Code is Meta's terminal coding agent, powered by the Muse Spark models, with subagents, MCP, hooks, skills and sandboxed …
Coding agent
OMP
Can Bölük / Stencil Labs, Inc.OMP (oh-my-pi) is the open-source coding agent harness by Can Bölük, a fork of Pi with LSP, debugger, sub-agents and access to …
Coding agent
OpenCode
AnomalyOpenCode is Anomaly's open-source AI harness for the terminal: a coding agent with free model choice across 75+ providers, MCP …
Coding agent
Pi
Earendil Inc.Pi is the open-source, minimalist coding agent harness from Earendil Inc. for the terminal – extensible via TypeScript, with …
Coding agent
Qoder CLI
Bright Zenith Private LimitedQoder CLI is Qoder's terminal coding agent for autonomous code changes, tests and commands, with quest, plan and headless modes …
Coding agent
Qwen Code
Alibaba (Qwen team)Qwen Code is Alibaba's open-source coding agent for terminal, IDE and desktop, working with Qwen, Claude, Gemini and …
From developer harness to enterprise harness
Coding harnesses are tools for developers. Business users, cost control and governance need two more layers.
CompanyGPT – the harness for business users
CompanyGPT gives any language model tools via MCP servers, integrations and n8n, company knowledge via companyRAG, roles from Entra ID or Keycloak, and audit logs – operated in your Azure environment or sovereignly on STACKIT.
Discover CompanyGPTAI Gateway – cost and governance for every agent
The AI Gateway bundles the model access of your tools and coding agents behind an OpenAI-compatible API: authentication via Entra ID, budgets and cost centres per team, user and agent, guardrails and complete logging.
View the AI GatewayFrequently Asked Questions
What is an AI harness?
An AI harness – also called an agent harness – is the software around a language model that makes it able to work. It drives the agent loop, provides tools and an execution environment, manages context and memory, enforces permissions, verifies results and coordinates multiple agents where needed. The model supplies the reasoning, the harness supplies the ability to act.
What is the difference between an AI model and an AI harness?
The model – such as Claude, GPT or Gemini – produces text and decisions. The harness – such as Claude Code, Codex or Pi – gives the model tools, context, permissions and an environment in which it can read files, run commands and check results. The same model can perform noticeably differently in different harnesses.
Which AI harnesses are there?
Widely used coding harnesses include Claude Code from Anthropic, Codex from OpenAI, the open-source agent Pi, GitHub Copilot, Cursor Agent, OpenCode, Gemini CLI, Cline and Kilo Code. In addition there are orchestrators such as herdr, which run and coordinate several of these agents side by side in one terminal.
What is herdr?
herdr is an open-source runtime for coding agents in the terminal. It does not replace the agents; it runs them in real terminals, shows each agent's state as working, blocked or idle, and lets agents drive each other through a CLI and socket API. Supported agents include Pi, Claude Code, Codex, OpenCode, GitHub Copilot CLI and Cursor Agent.
How do companies keep cost and data protection under control with coding agents?
What matters is the route by which the harness obtains the model. Many harnesses allow a custom endpoint. A central AI gateway bundles these requests, authenticates users, enforces budgets per team and agent, and logs usage. Source code and prompts go to the chosen model provider – that data flow belongs in your record of processing activities.
Is CompanyGPT a harness too?
Yes, but for a different audience. Coding harnesses such as Claude Code or Codex are tools for developers in the terminal and IDE. CompanyGPT is the harness for business users – with tools via MCP servers and n8n, company knowledge via companyRAG, roles from Entra ID or Keycloak, and audit logs. The two complement each other.
Which harness fits your team?
We advise on selecting, introducing and governing coding agents – from model access in the EU to cost control.