↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Harnesses/

Atomic Agent

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 5 minutes read (~964 words)
agent-curated 453 fully-ai-generated 556 harnesses 35 llama-cpp 1 llm=glm-5.3-flash 464 local-first 11 open-source 129 research-note 423
Table of Contents

Atomic Agent is an MIT local-first agent (TUI, CLI, and a Tauri desktop shell) that runs open-weight models through its own TurboQuant llama.cpp fork and drives your browser, files, shell, git, and MCP tools with the control loop and all state on your machine.

The bet is that small quantized models stay useful for long, tool-heavy work if the harness does the engineering: grammar-constrained (GBNF) tool calls, a byte-stable prompt prefix for KV-cache reuse, a bounded context tail, and externalized state, so a 9B model clears half of GAIA Level 1 through the loop rather than the model.

What it is
#

A developer-preview Node.js/TypeScript agent installed by curl script with self-update and a documented uninstall flow, released under MIT (v0.6.7 on 2026-10-07, plus a first desktop-v0.0.1 Tauri build the same day). One inference produces a JSON array of tool calls (grammar-constrained on a local llama-server), independent reads run in parallel, risky actions ask first, and long jobs continue past 25-step checkpoints to a 1,000-step or 2-hour ceiling. The maker maintains a TurboQuant llama.cpp fork (claimed up to roughly 6.4x KV-cache compression and 30-50% throughput gains from speculative decoding) and a managed mode that downloads and runs the backend for you. Surfaces go beyond the terminal: an HTTP API whose /v1/chat/completions maps one request to one full macro-turn, a Tauri sidecar for embedding, Telegram and Discord bots with approval buttons, and Fusion mode where one model plans and a pool of throwaway workers executes via fusion.delegate (workers cannot delegate, reach you, schedule tasks, or write memory). It imports skills, memory, sessions, and (opt-in) keys from Claude Code, Codex, Pi, Oh My Pi, Hermes, and OpenClaw.

Status
#

Active and quick-growing: 3,230 stars and 265 forks as of 2026-10-10, created 2026-04-21, pushed 2026-10-09, with a steady weekly release train and a developer-preview warning that APIs and behavior are still moving. The headline benchmark is self-run: on the public GAIA validation Level 1 split (53 tasks), Atomic Agent scored 69.8% (37/53) against Hermes at 58.5%, both driving the same local qwen-3.6-35b-a3b on one M4 Max with the same step budget, with per-task matrices, NDJSON traces, and logs published on the release tag; a model-scaling table shows 52.8% at 9B and 45.3% at 12B on the same split. No independent reproduction exists, and a Hacker News search finds no thread about the project as of 2026-10-10, so the adoption evidence is the star curve and the published artifacts, not independent technical discussion.

Star History Chart

Strengths
#

  • The most complete local-model engineering in this section: grammar-constrained tool calls, cache-friendly prompt stability, and a compression step are aimed squarely at the failure modes of small models on long tasks.
  • Fusion is a pragmatic subagent story: a planner delegates wide reads and drafts to disposable workers and merges the results, with the orchestrator barred from mutating tools.
  • The vendor-run benchmark publishes its artifacts (matrices, traces, logs on the release tag) and holds the model and hardware constant against a named rival, which is better evidence practice than most self-reports.
  • Full desktop tool surface with approval gating on every dangerous action, plus verify checks that never report an unchecked file as passing.

Cautions
#

  • Developer preview: the README says APIs, commands, config, and behavior are still moving, so any integration should pin a release.
  • The benchmark is vendor-run on one model, one machine, and one 53-task split, and the loser (Hermes) is also the comparison the vendor chose; treat 69.8% as a reproducible claim, not an independent one.
  • The ecosystem around the agent is broad consumer surface (Atomic Mail, Atomic Chat, Atomic Wallet, Sigma Browser, Atomic VPN on the vendor site), which is an odd neighborhood for a dev tool and worth factoring into trust.
  • Category adjacency: Telegram and Discord channels plus ClawHub skill installs overlap the assistant-runtimes family, so teams should decide whether they want a harness that answers your phone.

Pricing
#

Free, MIT, with no paid tier recorded as of 2026-10-10: local models cost nothing, and cloud models bill your own provider keys, including Claude Code and OpenAI Codex subscriptions driven through their signed-in CLIs.

Compared to
#

  • Ante: the other embedded-llama.cpp bet, roughly 15MB with the engine inside the binary; choose Ante for footprint, Atomic Agent for the fuller tool surface and Fusion.
  • Hermes: the rival it benchmarks against, a much larger assistant runtime with a learning loop; choose Hermes for channels and self-improvement, Atomic Agent for a local coding-and-desktop agent under your keys.
  • Nanocoder: the community-built local-first harness; Nanocoder has the wider documented local-engine list, Atomic Agent ships the deeper local-inference engineering.

Bottom line
#

Recommended for running actual work on small local models, where the harness-side engineering (GBNF calls, cache-stable prompts, externalized memory) is the difference between a demo and a workday. Not for anyone needing a stable API contract today (developer preview), and not for teams who want independently reproduced benchmark numbers before adopting.

Changes
#

  • 2026-10-10 - Created from the harnesses resolution pass (two of three category workers placed it here); the assistant-runtimes adjacency (chat channels, ClawHub skills) is recorded in the cautions.

See also
#

  • Ante - the minimal embedded-llama.cpp contrast
  • Hermes - the benchmarked rival across the category line
  • Nanocoder - the other community local-first harness
  • Harness Feature Matrix - the category comparison this note joins
  • FrontierHarness Eval - the independent harness benchmark to read before believing any vendor self-report

References
#