Autonomous coding agents

Autonomous Coding Agents

Use this umbrella when Hermes should hand work to a separate coding agent process rather than performing every edit itself. Pick the backend based on availability, task shape, and required interaction.

Selection guide

Agent	Best for	Default mode	Notes
Claude Code	Large feature work, refactors, codebase navigation, structured review	claude -p print mode	Strong non-interactive JSON/streaming modes; interactive mode needs PTY dialog handling.
Codex CLI	OpenAI-backed coding, PR reviews, parallel issue fixes	one-shot or background PTY	Useful with OpenAI-Codex OAuth and git worktrees.
OpenCode	Feature work, refactoring, PR review with open tooling	one-shot or background PTY	Resolve the binary path first; supports interactive sessions and resuming.

Prefer delegate_task for bounded subtasks inside the current Hermes turn. Use external CLIs for longer work, richer codebase autonomy, or tools/models not exposed through delegate_task.

Universal orchestration pattern

1. Inspect repo state and define the task with exact files, acceptance criteria, and constraints.
2. Choose one-shot mode for bounded work; choose tmux/PTY for multi-turn or interactive tools.
3. Run the agent from the correct workdir. Use git worktrees for parallel tasks that will edit files.
4. Monitor output; answer prompts explicitly; kill stuck sessions cleanly.
5. Verify the result yourself with tests, diffs, and file reads before reporting success.

Claude Code

Preferred for most non-interactive work:

claude -p "Implement X. Run tests. Summarize files changed."

Interactive mode requires a real PTY, usually tmux. First-run dialogs may require accepting workspace trust and the permissions warning. Do not run blind; capture the pane and verify that Claude is at the prompt or actively working.

Codex CLI

Use for OpenAI Codex-backed changes and PR reviews. Patterns:

codex "Review this PR for correctness and security"
# or background/PTY when it may ask questions

For parallel issue fixing, create separate git worktrees, launch one Codex process per worktree, then review/merge results centrally.

OpenCode

Resolve the binary first (opencode, opencode-ai, local install path, etc.). Use one-shot mode for simple tasks and a PTY for multi-turn sessions. Watch cost/session state and resume existing sessions when continuing prior work.

PR review workflow

• Fetch/check out the PR locally when possible.
• Ask the coding agent for findings, but verify every claimed bug against the diff and tests.
• Post only actionable comments; distinguish blockers from suggestions.

Pitfalls

• Never assume an external agent completed work because it said so; inspect the filesystem and test output.
• Do not launch multiple editors in the same worktree.
• PTY tools may hang on trust/permission prompts; handle dialogs before waiting for final output.
• Background processes should use Hermes background=true plus notify_on_complete=true for bounded runs, or tmux for interactive sessions.