Field Guide to Coding Agents

The coding-agent field has dozens of agents and very little rigorous documentation of what each one’s architecture actually is. Practitioners pick agents by leaderboard score, not by understanding the loop, tool set, context strategy, or termination heuristic underneath. This guide catalogs exactly that, for every agent Chimera replicates.

Two ground rules, inherited from the spec that produced these pages:

1. Replicated agents only. Every entry has a runnable Chimera primitive — either a full CLI replica (one of the seven codenames) or a loop-level agent style in chimera/agents/presets/agent_styles.py. The primitive path on each page is verified against a pinned commit.
2. Firsthand reading only. Each page was written from the upstream source, docs, or paper plus the Chimera replica’s code — never from secondhand summaries. Claims that could not be verified were omitted. Benchmark scores are deliberately absent; for controlled comparisons, use chimera bench-compare.

Taxonomy

Agent	Loop type	Edit format	Context strategy	Primary surface	Year	Chimera primitive
SWE-Agent	ReAct, one command per turn	line-range replacement, lint-guarded	history with old observations collapsed	autonomous batch runner	2024	swe_agent style
Aider	edit–apply–reflect chat loop	search/replace diff blocks	graph-ranked repo map	terminal chat	2023	aider style
Cline	Plan/Act dual-mode ReAct	diff-reviewed edits	tool-invocable compaction + checkpoints	IDE-embedded (VS Code)	2024	cline style
Claude Code	gather–act–verify tool loop	exact string replacement	auto-compaction (clear, then summarize)	terminal TUI	2025	chimera/mink/
Codex CLI	ReAct with sandbox-mediated escalation	apply_patch patch dialect	AGENTS.md chain + resumable sessions	terminal TUI + exec + IDE	2025	chimera/ferret/ + codex style
opencode	ReAct over a headless server	string-replace + patch + whole-file	auto-compact on overflow; LSP-checked edits	TUI client of a local HTTP server	2025	chimera/otter/
pi	ReAct, four operating modes	exact-string search/replace	JSONL session tree + summarizing compaction	terminal CLI (TUI/print/RPC/SDK)	2025	chimera/weasel/
little-coder	ReAct + 20 lifecycle extensions	search/replace with write-guard	per-turn budgeted skill injection	terminal, local llama.cpp models	2026	chimera/shrew/
Kimi CLI	checkpointed step loop	whole-file write + string replace	ratio-triggered auto-compaction	terminal REPL with shell-mode toggle	2025	chimera/stoat/
claw-code	streaming tool-call loop	tool-call file edits	resumable sessions; rerun resets	terminal REPL + one-shot	2026	chimera/badger/

Reading order

Start with the archetypes, then branch by the surface you care about:

1. SWE-Agent — the agent-computer interface archetype: bash as the primary tool, a ReAct loop, and the case for LM-tuned commands over raw shell.
2. Aider — the edit-format archetype: search/replace diff blocks, the repo map, and git as the undo stack.
3. Cline — the human-in-the-loop archetype: dual Plan/Act modes with approval-gated actions.
4. Then the terminal-CLI generation: Claude Code (TUI-first), Codex CLI (sandbox-first), opencode (server-first), pi (minimal harness), little-coder (small-model fit), Kimi CLI (shell-mode toggle), and claw-code (harness-rewrite discipline).

How these map to Chimera

The seven CLI postures are full replicas — each a shipping chimera subcommand documented in Inspirations with a per-codename parity matrix. The agent styles are loop-level replicas: compositions of Chimera’s own loop, tool, and prompt primitives that encode an upstream’s architecture in a few dozen lines (see the Build SWE-Agent in 60 lines tutorial for the pattern). Same model, same harness, different architecture — which is what makes them comparable under chimera bench-compare.