---
type: "post"
id: "yohei-nakajima-releases-babyagi-inside-babyagi-loop-compiled-f0e2b2"
slug: "yohei-nakajima-releases-babyagi-inside-babyagi-loop-compiled-f0e2b2"
title: "Yohei Nakajima releases babyagi-inside: BabyAGI loop compiled into Qwen weights"
description: "Yohei Nakajima’s babyagi-inside repository describes a bounded BabyAGI-style loop compiled into widened Qwen3.5-2B checkpoints."
retrieval_nugget: "Yohei Nakajima’s babyagi-inside repository describes a bounded BabyAGI-style loop compiled into widened Qwen3.5-2B checkpoints. In archived traces, the controller completes its fixed cycles and handles sequential arithmetic, but two structured tasks fail exact semantic scoring. The experiment matters because it makes an alternative to external agent orchestration concrete—and exposes its limits."
published_at: "2026-11-05"
updated_at: "2026-11-05"
record_date: "2026-11-05"
date_kind: "published_at"
topics: ["agent-loops","open-models","models"]
entities: ["Yohei Nakajima","babyagi-inside","Qwen3.5-2B","qwenFS"]
related_patterns: ["harness-architecture-outlives-model-choice"]
novelty_label: "new_fact"
what_changed: "A new project compiles a BabyAGI-style agent loop into Qwen weights, distinct from QwenPU’s vector-unit computation compiled into weights."
prior_refs: ["https://newruntime.com/posts/qwenpu-compiles-an-npu-into-qwen3-5-4b-077c19/"]
source_url: "https://github.com/yoheinakajima/babyagi-inside"
source_urls: ["https://github.com/yoheinakajima/babyagi-inside"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:yohei-nakajima-releases-babyagi-inside-babyagi-loop-compiled-f0e2b2"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/yohei-nakajima-releases-babyagi-inside-babyagi-loop-compiled-f0e2b2/","markdown":"https://newruntime.com/posts/yohei-nakajima-releases-babyagi-inside-babyagi-loop-compiled-f0e2b2.md","json":"https://newruntime.com/posts/yohei-nakajima-releases-babyagi-inside-babyagi-loop-compiled-f0e2b2.json"}
markdown_url: "https://newruntime.com/posts/yohei-nakajima-releases-babyagi-inside-babyagi-loop-compiled-f0e2b2.md"
---

# Yohei Nakajima releases babyagi-inside: BabyAGI loop compiled into Qwen weights

## Retrieval answer

Yohei Nakajima’s babyagi-inside repository describes a bounded BabyAGI-style loop compiled into widened Qwen3.5-2B checkpoints. In archived traces, the controller completes its fixed cycles and handles sequential arithmetic, but two structured tasks fail exact semantic scoring. The experiment matters because it makes an alternative to external agent orchestration concrete—and exposes its limits.

## Summary
Yohei Nakajima’s babyagi-inside repository describes a bounded BabyAGI-style loop compiled into widened Qwen3.5-2B checkpoints. In archived traces, the controller completes its fixed cycles and handles sequential arithmetic, but two structured tasks fail exact semantic scoring. The experiment matters because it makes an alternative to external agent orchestration concrete—and exposes its limits.

## What happened
Yohei Nakajima published babyagi-inside as an open-source systems experiment. The repository says it compiles a bounded BabyAGI-style agent loop into a widened Qwen3.5-2B checkpoint. Its central change is where the control logic lives: the host provides one objective, then samples model output until it receives HALT. According to the repository, the host does not parse phases, execute tools, inject follow-up prompts, repair outputs, or manage a task list after that initial objective.

The checkpoint’s widened residual stream is partitioned among three functions. Dimensions 0–2047 are Qwen’s native model state; 2048–3071 are assigned to qwenFS, the compiled filesystem state and operations; and 3072–3208 are for the agent controller and bounded-agent scratch state. The cycle count is fixed when a checkpoint is constructed. The verified set uses separate checkpoints for 2, 3, 5, and 8 cycles. The model writes the plan, reasoning, and payloads, while the compiled controller advances a fixed task pointer and qwenFS persists the ledger and results.

The repository reports ten completed live compiled-checkpoint diagnostics. All reach HALT with the expected cycle count and pass checks for controller behavior, qwenFS reads and writes, and disjoint token accounting. Eight sequential arithmetic chains match their frozen expected results across the tested cycle counts. Two eight-cycle structured-operation traces complete mechanically, but their exact semantic scorer results fail. The authors frame the outcome as a systems demonstration, not evidence of general agency; the structured-planning bottleneck remains unresolved.

The weights are not shipped in the repository. A builder must provide a local Qwen3.5-2B model matching the recorded SHA-256 provenance and run the documented build script. The README estimates each resulting checkpoint at approximately 8.8 GB and asks users to have roughly 10 GB free per checkpoint. Reproduction of the live traces requires Apple Silicon with MPS, the matching base weights, and `uv`. The repository does not preserve verified build-time or peak-memory measurements: both are marked TODO, and it advises building one checkpoint at a time until those costs are measured. It records a qwenFS-derived compiler source revision and an upstream qwenFS baseline, but says the upstream Qwen model revision ID was not recorded.

## What the archive adds
This is a new fact in the broader thread of putting more capability into Qwen weights. Earlier notes cover [QwenPU compiling a vector unit into Qwen3.5-4B](https://newruntime.com/posts/qwenpu-compiles-an-npu-into-qwen3-5-4b-077c19/) and [Qwopus3.8-27B-Flash-V2, a Qwen fine-tune optimized for agent loops](https://newruntime.com/posts/qwopus3-8-27b-flash-v2-is-an-apache-526d19/). babyagi-inside is distinct: it embeds bounded controller and filesystem state, not a vector unit or a post-training optimization claim.

## Trend signal
This experiment complicates the claim that the harness is a more durable boundary than the model. The repository moves a bounded controller and persistence mechanism into a checkpoint, but its cycle count is fixed at construction and the structured-task results are limited. It therefore tests—rather than overturns—the [harness architecture outlives model choice](https://newruntime.com/patterns/harness-architecture-outlives-model-choice/) pattern: teams still need to compare the operational flexibility of an external harness with the constraints of in-weight control.

## Why it matters
For teams building agents, the useful question is not simply whether control logic can be represented inside model weights. It is what that choice removes from the host, and what new constraints it introduces. Here, the host loop becomes unusually small: prefill an objective and sample until HALT. That is a meaningful systems property, especially for people investigating how much orchestration can be compiled into a model-facing runtime.

But the repository’s own evidence keeps the claim narrow. A predictable control trace is not the same as a correct task result. The arithmetic results show that the compiled loop can support a bounded class of sequential tasks; the two structured failures show that completing the mechanical protocol does not ensure semantic success. Fixed cycle counts also mean that a builder cannot treat this as an ordinary flexible agent scheduler. And rebuilding substantial checkpoints locally, with unmeasured time and memory requirements, makes experimentation costly and hardware-dependent.

The practical lesson is to separate controller verification from task evaluation. If a team explores in-weight agents, it should score both: whether state transitions, reads, writes, and stopping behavior are correct, and whether the requested work is actually correct. This repository gives builders an example of both kinds of evidence, including failures, rather than presenting a successful trace as proof of general autonomy.

## Limits and uncertainty
The supplied source documents one project’s implementation and ten archived live traces; it does not establish broad performance, general agency, or superiority to an external scheduler. Eight arithmetic successes and two structured-task semantic failures are a small, bounded evaluation, not a general benchmark. The repository does not report verified build duration or peak memory, and the upstream Qwen revision ID was not recorded. The source also does not establish that the approach is cheaper, faster, safer, or more reliable than a conventional harness, nor does it provide a ready-to-download compiled checkpoint.

## Skill to build now
A builder can use the repository’s split between mechanical diagnostics and semantic scoring as a review template this week. For an existing agent loop, write separate checks for state transitions, persistence, and stop conditions, then score the requested task outcome independently. Keep failures visible rather than treating a clean HALT or completed trace as evidence that the agent did the job correctly.

## Reader takeaway
babyagi-inside demonstrates that a small, bounded agent controller can be compiled into widened Qwen weights, reducing the host’s job to supplying an objective and sampling until HALT. Its traces verify the mechanics more strongly than its general usefulness: arithmetic succeeds, structured planning does not. Treat it as an inspectable architecture experiment, not a drop-in autonomous agent.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Agent Loops](https://newruntime.com/topics/agent-loops.md): Explore the agent-loops topic hub.
- topic hub: [Open Models](https://newruntime.com/topics/open-models.md): Explore the open-models topic hub.
- topic hub: [Models](https://newruntime.com/topics/models.md): Explore the models topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)