docs: comprehensive agent harness reference and selection guide #2372

Background

The ecosystem page lists three harnesses as "some examples" and links to their directories. With 34 agent harnesses now in the repo — 20 general-purpose and 14 benchmark-specific — there is no single place where a user can understand what is available, how to choose between them, or how to configure and run one. This creates friction for anyone evaluating a new model or building an environment.

We should expand the Configure Agents tab so to include relevant information on Agents with Gym.

Problems

• No complete list of available harnesses in the docs
• No guidance on when to use one harness over another (e.g. simple_agent vs claude_code_agent vs hermes_agent)
• No explanation of the distinction between general-purpose harnesses and benchmark-specific agents
• No documentation for harnesses added in 0.5.0: codex_agent, kilocode_agent, anyswe_agent, remote_agent
• The ecosystem page references OpenHands as a standalone harness but it is accessed through swe_agents
• Configuration options (model wiring, ModelServerRef, sandbox attachment, skill loading) are undocumented at the harness level

Proposed Documentation

1. Agent Harnesses reference page — a table of all general-purpose harnesses with: name, the external tool or framework it wraps, primary use case, and a link to its directory README. Clearly separates general-purpose from benchmark-specific agents.
2. Choosing an agent harness — a short decision guide:

• Single-turn evaluation with no tools → simple_agent
• Coding / SWE tasks → swe_agents (OpenHands or OpenCode), mini_swe_agent, anyswe_agent
• Agentic coding with external CLI → claude_code_agent, codex_agent, kilocode_agent, opencode_agent, openclaw_agent
• Multi-step reasoning with LangGraph patterns → langgraph_agent
• Driving an agent you host yourself → remote_agent
• RL multi-turn with simulated user → gymnasium_agent

3. Configuring a harness — cover the common configuration surface shared across harnesses:

• Wiring a model server via ModelServerRef
• Attaching a sandbox provider
• Loading agent skills from .agents/skills/
• Per-request policy endpoint override (for RL training integration)

4. Ecosystem page update — replace the three-item "some examples" list with a link to the new reference page and a one-line summary.

Out of scope

Benchmark-specific agents (e.g. stirrup_agent, scicode_agent) are documented alongside their benchmarks and do not need a separate harness page.