惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

量子位
雷峰网
雷峰网
博客园 - 三生石上(FineUI控件)
月光博客
月光博客
有赞技术团队
有赞技术团队
阮一峰的网络日志
阮一峰的网络日志
Last Week in AI
Last Week in AI
G
Google Developers Blog
腾讯CDC
B
Blog
Microsoft Azure Blog
Microsoft Azure Blog
酷 壳 – CoolShell
酷 壳 – CoolShell
Microsoft Security Blog
Microsoft Security Blog
人人都是产品经理
人人都是产品经理
博客园_首页
T
Tailwind CSS Blog
C
Check Point Blog
博客园 - 【当耐特】
MongoDB | Blog
MongoDB | Blog
A
About on SuperTechFans
Y
Y Combinator Blog
L
LangChain Blog
Engineering at Meta
Engineering at Meta
GbyAI
GbyAI

Hacker News - Newest: "AI"

AI can't read an investor deck AI as an attorney? Student uses ChatGPT, Gemini to sue UW over alleged racial discrimination Hacking MCP Servers in AI Systems – The Rug Pull: Tool Changes After Approval GitHub - MeepCastana/KubeezCut: Free Web based video editor Can AI judge journalism? A Thiel-backed startup says yes, even if it risks chilling whistleblowers Coming soon: 10 Things That Matter in AI Right Now DARPA built an AI to fact-check enemy weapons claims What explains heterogeneity in AI adoption? When AI Meets Muscle: Context-Aware Electrical Stimulation Promises a New Way to Guide Human Movements - Department of Computer Science AI Changed How We Build. It Did Not Change What Matters. Linux rules on using AI-generated code - Copilot is OK, but humans must take 'full responsibility for the… Meta spins up AI version of Mark Zuckerberg to engage with employees Code Mode: Let Your AI Write Programs, Not Just Call Tools | TanStack Blog GitHub - Delavalom/graft: Go framework for building AI agents. Type-safe tools, multi-provider (OpenAI, Anthropic, Gemini, Bedrock), zero vendor SDKs. India's TCS tops estimates, says new AI models did not dent services demand Gen Z's fading AI hype Strong feeling: we are in a folded AI reality GitHub - machinarii/total-recall-catalog: A reference catalog of latest knowledge retrieval, memory & RAG systems GitHub - mensfeld/code-on-incus: Give each AI agent its own isolated machine with root, Docker, and systemd. Active defense detects and stops threats automatically.. Quantization, LoRA, and the 8% Problem: Benchmarking Local LLMs for Production AI Iran war: We spoke to the man making Lego-style AI videos that experts say are powerful propaganda Powell, Bessent discussed Anthropic's Mythos AI cyber threat with major U.S. banks GitHub - immartian/bellamem: Persistent belief-graph memory for AI agents. Retrieves decisive context by importance — not recency, not RAG, not /compact. recursive-mode: The Repo-Native Operating System for AI Engineering After the attack on Sam Altman's home, will AI CEO's go on the offensive? The biggest advance in AI since the LLM Opus 4.6 vs GPT 5.4 One Prompt Unity World Generation Test “AI polls” are fake polls Client Challenge Can AI be a 'child of God'? Inside Anthropic's meeting with Christian leaders
GitHub - tamerh/enju: Coordinating Humans, AI Agents, and...
timurg · 2026-05-27 · via Hacker News - Newest: "AI"

Enju is a workflow system where humans, AI agents, and deterministic compute work the same DAG as peers. The unit of work is a task — something any of them can answer, review, vote on, or compute. The graph is live: a task can spawn more tasks while a run is in flight, so a review that returns request_changes drops a revision task back into the graph with its feedback already attached, and the work cycles until it's approved.

What makes this work is where the lines are drawn. Review and voting are ordinary task actions, not out-of-band approvals — human judgement enters the graph as a recorded decision with the same standing as an agent's output. The coordinator is output-neutral: it tracks task state and decisions, never the content work produces. Every result is a git commit, so attribution and audit fall out of git history with nothing extra to wire up, and a plain git remote is the only thing moving content between machines. Enju ships as a single binary that speaks MCP, a CLI, and a web UI.

A PRISMA systematic-review workflow in Enju: compute, AI agents, and human review gates as peers on one DAG

A real Enju workflow — a PRISMA systematic review — where deterministic compute (teal), AI agents (blue), and human review gates (orange) are peers on one graph.
Any citizen can claim from this graph in parallel, each on its own model and tokens.

What a workflow looks like

A workflow is a DAG of tasks, written as YAML and committed to your repo. Here an agent drafts a report and a human gates it — two tasks, two different kinds of citizen, one graph:

name: My First Workflow

agents:
  - name: writer
    handler: claude
    model: claude-sonnet-4-6

tasks:
  - id: write_report
    action: answer            # an agent (or a human) produces work
    assign_to: writer
    writes: [report.md]
    prompt: Write a short report on solar-energy adoption to report.md.

  - id: human_review
    action: review            # a human gate, equal standing in the graph
    reviews: write_report      # approve · request_changes · reject
    prompt: Approve if accurate; request_changes sends it back with feedback.
enju go enju.yaml --auto-agents

The agent claims write_report, runs its model, and commits report.md; human_review then waits in your inbox. Every step is a commit on the run's branch. → full walkthrough in the quickstart.

How it fits together

Enju architecture

The coordinator holds the task DAG and its lifecycle (pending → ready → claimed → running → review → done, with a revise loop) plus the state and events databases — but no produced content. Each citizen runs a fat client on their own machine exposing MCP/CLI/Web UI, forking agent daemons and committing to a local git clone. Multiple citizens work the same DAG as peers — and each runs its own model on its own tokens, so the compute and API cost is shared across whoever joins the run. Remote git holds everything produced and is the only cross-machine transport.

The task model

One primitive, interchangeable executors: a task's action selects whether a human, an LLM agent, or a script runs it — all the same kind of node. Edges carry typed data, for_each fans a task (or a whole run) out into parallel iterations, and a review verdict can approve, fail, or cycle the work back with feedback — every attempt kept as a commit.

Quick install

curl -fsSL https://raw.githubusercontent.com/tamerh/enju/main/install.sh | sh

Installs enju to ~/.local/bin/enju (no sudo). Add it to your PATH if it isn't already:

echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.bashrc   # or ~/.zshrc

Verify with enju --version.

Other platforms or specific versions: download a binary from the releases page and put it on your PATH.

Examples

Three reference workflows — clone, install, run:

  • mustache-engine-enju — build a Mustache template engine from spec. Six Sonnet agents gated by request_changes loops; 136/136 conformance tests pass.
  • prisma-review-enju — PRISMA systematic review of FMT-for-rCDI RCTs. Four Sonnet agents + two human review gates produce a 14-RCT synthesis.
  • nanopore-assembly-enju — ONT phage-genome assembly. Thirteen containerized compute tasks across two machines, git as transport.

Docs

See docs/getting started, guides, reference, or how it works.

For the design and motivation, see the preprint: sugi.bio/enju.

License

MIT — see LICENSE.