agent-body

English · [中文](README.zh-CN.md)

Agent‑Body

Load the tools your task needs. Give your agent memory, reflexes, and a recovery loop.

A plugin layer for [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) that routes intent to capabilities and keeps unused tool definitions out of the model’s context.

26 organ identities · 84.7% of tool-schema tokens gated cold-start · 200+ offline assertions · MIT

Your agent may have hundreds of tools. A single task usually needs only a few. Agent-Body adds a runtime layer that decides which capabilities to expose, remembers outcomes, and attributes failures before retrying.

[Watch the recorded demo](https://1420079678-ctrl.github.io/agent-body/en.html) · [Read the benchmark](benchmarks/results/REPORT.md) · [Install in DeepSeek Harness](#install-in-one-line) · [中文](README.zh-CN.md)

| Problem | What Agent-Body does | Evidence you can inspect |

| --- | --- | --- |

| Tool definitions consume context on every turn | Gates schemas by task intent | 48-command offline benchmark: 55,154 → 8,433 estimated schema tokens on average |

| Routing adds another model decision | Routes commands through deterministic intent rules | Offline end-to-end demo; routing needs no model call |

| Failures trigger unhelpful retries | Classifies causes and selects a remedy; argument errors are not retried | Recovery trace in the recorded runtime output |

| Useful outcomes disappear between sessions | Adds persistent memory and idle consolidation through the Cortex organ | [Architecture](ARCHITECTURE.md) and included plugin source |

The 84.71% reduction is a cold-start tool-schema estimate, not a total-prompt or billing claim. The benchmark exposes 86.67% of task-declared required capabilities on the first turn; six are capped and must be retrieved via body_call. A recorded installation with history shows 58% schema gating. [See the full trade-off and methodology](benchmarks/results/REPORT.md).

The body metaphor maps to concrete components: plugins are organs, intent routing is the nerve bus, deterministic event handlers are reflexes, and the heartbeat coordinates state. This is an early-stage, Windows-first project. Host integration requires DeepSeek Harness; the offline core demo below runs independently.

Try it in 30 seconds — no install, no host, no API key. The core imports nothing outside Node built-ins, so a fresh

clone runs the real end-to-end chain offline:

git clone https://github.com/1420079678-ctrl/agent-body && cd agent-body
npm run demo     # command → impulse → dispatch → execute → attribute → reflex fires
npm run check    # the gate CI runs: constant tables, catalog, tests, benchmark — all offline

Here is that same body while it runs. Not a mockup — the frames are real tool output from a real install, rendered by

tools/make-demo-gif.py out of [docs/demo/captured.json](docs/demo/captured.json), which records the raw output. The

renderer refuses to draw a number that is not in that file:

Read it left to right: anatomy and vitals, then the token ledger, then a Chinese command becoming a nerve impulse

that innervates four organs — each told which capability to fire, with zero model calls spent deciding the route —

and finally a wound attributed to arg_error, prescribed do-not-retry, and closed by re-check.

Already running DeepSeek Harness? One command installs the body kernel, the memory organ and the context engine:

$rel = "https://cdn.jsdelivr.net/gh/1420079678-ctrl/agent-body@v0.1.2/dist"
dsh plugin --profile web add "$rel/dsh-external-dsh-organism-0.1.1.tgz" "$rel/dsh-external-dsh-cortex-0.1.1.tgz" "$rel/dsh-external-dsh-zero-residence-0.1.0.tgz"

Restart the harness and body_status lists the organs. The claim you can check for yourself: the tool-schema block

of the prompt drops 84.7% on a cold start across 48 representative commands, and npm run bench:check fails the

build if that number drifts. The scope is stated wherever the number appears — tool-schema tokens only, not the whole

prompt. On a body that has been running for a while it is 58% (127 of 332 visible); that figure, and why it

declines, are in [Token economy](#token-economy).

⭐ [Star the repository](https://github.com/1420079678-ctrl/agent-body/stargazers) if you want it to keep tracking

the host closely — it is a one-person project. A star helps other developers discover it; a reproducible issue or a new organ helps improve it.

---

[Architecture](ARCHITECTURE.md) · [Organ Catalog](catalog/organs.json) · [Benchmark](benchmarks/results/REPORT.md) · [Roadmap](ROADMAP.md) · [中文文档](README.zh-CN.md)

[Official DSH discussion — Show Your Plugins!](https://github.com/deepseek-ai/deepseek-harness/discussions/7555) · the channel the harness CONTRIBUTING points plugin authors to

v0.1.1 · MIT · Windows-first (Node 22.19 / 24) · 26 organ identities in the catalog, realised by 24 plugin packages in this repository · [release notes](https://github.com/1420079678-ctrl/agent-body/releases/tag/v0.1.1)

▶ [Open the recorded demo](https://1420079678-ctrl.github.io/agent-body/en.html) ([中文](https://1420079678-ctrl.github.io/agent-body/)) — a recorded replay of a real install: the heartbeat, the organs, the pulse stream, the healing ledger and the token gate, with nothing installed. It is generated from the runtime files (vitals.json, bloodstream.json, pulse.jsonl), not retyped from screenshots.

Contents · [Where it's listed](#where-its-listed) · [Install in one line](#install-in-one-line) · [Why this exists](#why-this-exists) · [What makes it different](#what-makes-it-different) · [The five biological layers](#the-five-biological-layers) · [Architecture at a glance](#architecture-at-a-glance) · [Organ catalog](#organ-catalog) · [Quick start](#quick-start) · [Write your own organ](#write-your-own-organ) · [Verify it yourself](#verify-i