The LLM Tool Universe

This document surveys LLM tools through the lens of the interaction pattern.

The lens asks four questions about a tool:

• Sync or async?
• How much human attention does it demand?
• What is the review surface?
• How does it integrate with git?

Terminal Code Agents

• Aider is a terminal code agent that edits files and commits with git.
• Aider is git-native.
• Every Aider change is a diff to review.
• Aider refreshes its repo map and warns about externally edited files.
• Aider mitigates stale context but does not isolate it.
• Claude Code is a terminal agent with background modes.
• Background agents run and notify on completion.

IDE-Integrated Agents

Cursor and GitHub Copilot embed agents in the editor.

Copilot has a background-agents pattern built on git worktrees.

Worktree Multiplexers

• dmux runs parallel coding agents, each in its own worktree.
• gwt runs AI coding sessions per worktree.
• gtr is a git worktree runner.
• rove runs coding agents on parallel tasks with isolated worktrees.
• These tools exist because the worktree pattern is real practice.

Harness Environments

• Agent harnesses orchestrate LLM work at scale.
• Harnesses dispatch work to subagents and notify on completion.
• Harnesses can serialize access so the human and the model never write concurrently.
• The environment hosting this conversation is such a harness.

CI and PR-Based Agents

• OpenHands and similar agents take an issue and produce a pull request.
• The PR is the envelope.
• The platform review and merge queue are the async mechanics.

Notebooks

• Jupyter notebooks are a thinking medium.
• Notebooks keep state, re-run cells, and mix text with results.
• Notebooks have hidden state and noisy diffs.

Plain API and Scripting

• The model API can be driven directly.
• A queue of prompts can run as a batch job.
• Results land in files for later review.
• This is the email pattern with minimal tooling.

Mapping to This Project

• Aider fits the review-surface pattern.
• A worktree gives Aider isolation.
• An episode file is the work order.
• A record file is the outcome.

Sources

• June 2025 Coding Agent Report: evaluation of 15 coding agents. https://github.com/The-Focus-AI/june-2025-coding-agent-report
• ai-agent-benchmark: comparison of 80+ agents. https://github.com/murataslan1/ai-agent-benchmark
• Claude Code background agents. https://github.com/marc-shade/agentic-system-oss/blob/master/intelligent-agents/CLAUDE_CODE_BACKGROUND_AGENTS.md
• Background coding agents with git worktrees. https://crazyrouter.com/en/blog/background-coding-agents-without-lock-in-git-worktrees
• dmux. https://github.com/sterlingchapman/dmux
• gwt. https://github.com/slowestmonkey/gwt
• gtr. https://landscape.jimmysong.io/zh/projects/git-worktree-runner/
• rove. https://github.com/Sma1lboy/rove
• aider-desk worktree isolation docs. https://deepwiki.com/hotovo/aider-desk/2.5-git-worktrees-and-isolation
• Parallel AI coding context collisions. https://1devtool.com/blog/context-collisions-parallel-ai-coding
• OpenHands. https://github.com/OpenHands/OpenHands

Status

This document is a living survey.

Verify tool names and features before relying on them.