Agents and Harnesses
A harness is the runtime that gives a model a loop, tools, memory, and an environment in which to act. It determines how long an agent can run, what it can access, and how its work can be observed and controlled.
Modern harnesses
| Name | Type | Primary Focus | Notes / Key Traits | Stars (approx, Oct 2026) |
|---|---|---|---|---|
| goose | Both | General-purpose local agent | Own harness + agent; MCP-native; desktop + CLI; AAIF | ~55k |
| Hermes Agent | Both | Always-on personal agent | Own runtime; persistent memory + skill creation; messaging gateways | ~250k+ |
| Agent Zero | Both | Full computer-use agent | Own Docker Linux desktop + agent; high transparency | ~19.5k |
| OpenClaw | Both | Personal assistant (chat apps) | Own gateway/runtime; messaging-first | Very high (~390k+) |
| OpenHands | Both | Coding / software engineering | Own multi-agent runtime; web UI + desktop | ~89k |
| OpenCode | Both | Terminal coding agent | Own agent loop; model-agnostic | High |
| Codex (OpenAI) | Both | Coding agent suite | Cloud + CLI + IDE; proprietary harness | N/A (product) |
| Claude Code | Both | Coding agent | Anthropic harness + agent; subscription-based | N/A (product) |
| Cline | Both | IDE coding agent | VS Code / JetBrains extension with own loop | — |
| Aider | Both | Terminal coding agent | Own agent loop; git-aware | — |
| Eigent | Both | Multi-agent desktop | Parallel agents on one task | ~15.5k |
Compare harnesses by their philosophy, strengths, limitations, deployment style, and fit for the task.
Supporting infrastructure
- MCP (Model Context Protocol) connects agents to tools and services.
- Sandboxed execution environments such as Docker, Daytona, Modal, E2B, and browser infrastructure help isolate agent actions.
- Persistent memory and agent-written skills support longer-running work.
- SpyNEL orchestrates existing coding harnesses.
- DOX / AGENTS.md hierarchies give agents maintainable project context.