ML Intern: a HuggingChat agent that trains, evaluates and publishes models from a prompt

ML Intern is an agent mode in HuggingChat that takes a plain-text brief, plans the training work, asks for budget approval before spending anything, runs a smoke test, then trains, evaluates, and publishes a model to Hugging Face. One author used it to build six models over a few days — fine-tuned VLMs, LoRAs, and distilled models — for a total compute cost of about $103.

• Agent enforces a hard spending cap per project; it asks permission before any paid job runs.
• Prompt structure that works: state verified facts upfront, request a pre-training baseline, require a smoke test with a pass/fail check before the full run.
• Six projects ranged from $1.90 (citrus disease VLM fine-tune) to $37 (two-run model distillation); each ended as a published model with an evaluation in the model card.
• Available now in HuggingChat — switch on ML-intern mode and paste a prompt.

📖 Read the full article
🔗 Source

#AIAutomation #MLAutomation #HuggingFace #LLMOps #AgentWorkflow

ARTEX: multi-agent pentest system with dual-graph state and persistent planner todolist

ARTEX is an AGPL-3.0 autonomous penetration testing system (Go backend, Next.js frontend) built around a planner-worker multi-agent architecture. Its core contribution for agent builders is a dual-graph state design: a global asset graph and a per-task exploration graph linked by anchors, plus a planner that keeps a shared todolist across stateless LLM invocations to drive multi-step dependency chains.

• Planner-worker split: planner dispatches intents from a persistent cross-invocation todolist; N workers each claim and execute one intent, then write facts back to both graphs.
• Workers can search other workers' raw execution traces to reuse intermediate observations that were never formally written as facts.
• Dual-graph state: global Asset Graph (truth store across tasks) + per-task Exploration Graph (lineage of goals, intents, facts, findings), joined via anchor records.
• Supports MCP servers, human-in-the-loop chat, tool approval/intercept gates; requires PostgreSQL and an Anthropic or OpenAI API key.

📖 Read the full article
🔗 Source

#AIAutomation #MultiAgent #AgentArchitecture #LLMOps #MCP

AI Automation Roundup, Oct 5-12, 2026: Agents Build and Search

This week's roundup covers ten stories on AI agents that build, search and operate software, from model training to on-screen guidance.

📖 Read the full roundup

#AIAutomation #Weekly

VOYGR called 1,000+ SF businesses with an AI agent that disclosed itself — only 1 in 3 stayed on

VOYGR (YC W26) used their PlaceCall API — which lets AI agents make real outbound phone calls — to call over 1,000 San Francisco businesses during SF Tech Week. The agent disclosed it was an AI on every call and asked one question. Only 1 in 3 businesses kept talking; most hung up within 5 seconds.

• PlaceCall is VOYGR's API for building AI agents that make real outbound phone calls.
• Engagement varied sharply by business type: vets 79%, dentists 61%, bars and dry cleaners 20%.
• About 1 in 25 businesses already has an AI answering, producing AI-to-AI call scenarios.
• Upfront disclosure of AI identity was used on every call — a real design variable affecting completion rates.

📖 Read the full article
🔗 Source

#AIAutomation #VoiceAgents #AICallers #AgentDesign #PlaceCall

n8n workflow parses CV PDFs into a structured four-tab Google Sheet via easybits Extractor

This n8n template accepts CV PDFs through a built-in web form, runs them through the easybits Extractor community node to pull 10 structured fields, then fans the results across four Google Sheet tabs: experience, education, skills, and summary. The sheet acts as a reusable data foundation you can wire into screening, tracking, or follow-up email workflows later without reworking the extraction layer.

• Extracts 10 fields per CV: roles, bullets, education, skills, contact info, languages, and portfolio links.
• A Code node fans out the Extractor response into four parallel Google Sheets append branches.
• easybits Extractor is a verified n8n community node; easybits offers a free account tier that covers the 10-field limit used here.
• Designed as a composable foundation — screening, ATS sync, or email follow-up can be added on top without redoing extraction.

📖 Read the full article
🔗 Source

#AIAutomation #n8n #DocumentExtraction #CVParsing #GoogleSheets

n8n workflow automates long-tail keyword research from a single seed keyword

This n8n workflow takes a seed keyword from Airtable and fans it out through three parallel research paths via YepAPI: related keywords, keyword ideas, and autocomplete suggestions. The result is a structured list of keyword candidates without manual lookups. The workflow is importable as a .json file and is MIT licensed.

• Triggers from a seed keyword stored in Airtable, runs three YepAPI research paths in parallel.
• Covers related keywords, keyword ideas, and autocomplete suggestions in one automated pass.
• Part of a broader MIT-licensed n8n workflow collection covering AI, CRM, messaging, and more.
• Import by dropping the .json into the n8n editor; requires your own API keys and credentials.

📖 Read the full article
🔗 Source

#AIAutomation #n8n #SEOAutomation #LongTailKeywords #WorkflowAutomation

REA gives coding agents tools to reverse engineer apps, binaries and browser scripts

REA connects to your coding agent and lets it inspect running software - returning assembly, decompiled code, and call graphs - so the agent can explain what a program does in plain English. It works on native binaries, JavaScript/Electron apps, and live browser scripts. Setup runs via a single npx command.

• Agents can inspect native binaries, Electron ASAR archives, or live browser scripts via a local debugging connection.
• REA returns raw assembly, decompiled code, and call relationships; the agent then interprets and explains them.
• Installed with `npx rea-agents@latest setup` and wired directly into your existing coding agent.
• Practical use cases shown: recovering a game's speed rule, rebuilding a calculator's % button logic from the installed exe.

📖 Read the full article
🔗 Source

#AIAutomation #CodingAgents #ReverseEngineering #AgentTools #LLMOps

bigarrow: a macOS tool for AI agents to point at UI elements when a human must act

When an AI agent hits a step only a human can complete — an OAuth prompt, a 2FA field, a permission dialog — it can now draw a labeled arrow, box, or ring directly on screen pointing at the exact element. bigarrow is a macOS CLI tool and skill for Claude Code and Codex, MIT licensed, with no daemon, no telemetry, and no screen-recording permission required for basic use.

• Agents call `bigarrow point` or `bigarrow start` targeting UI elements by label, coordinate, or Accessibility ID.
• Overlays are click-through everywhere except the sign; keyboard focus never moves to bigarrow.
• `--say` reads the sign aloud; `{{value}}` in sign text becomes a one-click copy button for codes or URLs.
• Install via Homebrew; `bigarrow install-skill` registers the skill with Claude Code and Codex automatically.

📖 Read the full article
🔗 Source

#AIAutomation #AIAgents #MacOSAutomation #HumanInTheLoop #ClaudeCode

GoLive: agent skill that deploys AI-built apps to production on your own accounts

GoLive is a MIT-licensed Agent Skill and zero-dependency Node CLI that takes agent-built apps from code to live infrastructure. It follows a detect → plan → approve → apply → verify loop across hosting, databases, DNS, email, and payments — all on your own provider accounts, with no GoLive backend or telemetry. Currently early alpha (0.1.0-alpha.8), verified with Claude Code and Codex.

• Covers Vercel/Netlify hosting, Supabase/Neon databases, DNS (Porkbun/GoDaddy), transactional email (Resend), and Stripe test-mode payments.
• Approve-first: no write reaches a provider account without a reviewed plan ID and explicit confirmation flags.
• Teardown command removes only resources GoLive itself created, naming anything it cannot remove for manual cleanup.
• MIT licensed, no external account required; install via npx, npm, or ClawHub; needs Node.js 20+.

📖 Read the full article
🔗 Source

#AIAutomation #AgentSkills #ClaudeCode #AIDeployment #LLMOps

Intent-Router: compile vague requests into typed IntentSpec contracts before your agent acts

Intent-Router is a prompt-only skill (MIT, Python) that converts a vague request into a typed IntentSpec contract. It probes your repo, docs, or ticket system to resolve unknowns before asking a human anything, then asks exactly one focused question for decisions that can't be looked up. The resulting YAML contract persists across sessions and is checked against the delivered work.

• Probes code, config, docs, and ticket data to resolve unknowns automatically — user is not interrupted.
• Asks one question only when the answer is a human judgment call or an irreversible decision.
• Saves a typed IntentSpec YAML per intent; later agent sessions and teammates reuse it instead of starting from zero.
• Distinguishes underspecified requests from degraded probes — preventing silent routing failures that look like clarification needs.

📖 Read the full article
🔗 Source

#AIAutomation #LLMRouting #AgentWorkflows #ContextEngineering #AIAgents

How Cockroach Labs structured a multi-agent coding pipeline around a teaching hospital

Cockroach Labs built MOLT Sinai, a GitHub Actions-based agent pipeline where distinct roles — Triage Nurses, Fellows, Review Attendings, Discharge Nurses — enforce mandatory planning before code, structured escalation, and post-merge audits. Over five months it processed more than a million lines of code with only 7 reverts. IBM Db2 migration support was shipped in under two days at $4,172 in tokens; the equivalent Oracle work done by humans took 9 months and ~$160K.

• Each agent role is defined by a Markdown skill file; GitHub issue labels drive state transitions between pipeline stages.
• Hard rules: no code written before a plan is reviewed by a separate Review Attending agent instance; agents must not modify tests to make them pass.
• Stuck agents file a structured I-PASS handoff and escalate rather than retrying; human decisions that generalize are stored in an append-only precedent log agents must consult first.
• Total cost over 5 months: ~$135K in Claude tokens, averaging ~$84 per GitHub issue; 85% of commits to the pipeline repo were authored by the pipeline itself.

📖 Read the full article
🔗 Source

#AIAutomation #AgentWorkflows #LLMOps #CodingAgents #MultiAgentSystems

Tiered model pipeline searches 4.35 million archive pages overnight, surfaces unrecorded history

A software engineer built an agentic research pipeline over 4.35 million handwritten Dutch East India Company pages and historical newspapers. The core trick: a cheap, fast filter model screens candidates first so the heavier model only reads what passes. The pipeline found an unrecorded 1812 meteorite fall, three undocumented Javan rhino shipments, and three volcanic eruptions absent from the Smithsonian's global list.

• Tiered model design: a small fast model (Jev) gates 59,000 passages before Claude Haiku reads; 59k elephant mentions cost ~$3 USD.
• Semantic embeddings on 5.7M passages beat keyword search on noisy OCR and archaic spelling — found 'meteorite' without the word ever appearing.
• Validation step built in: pipeline must rediscover known events (Laki 1783, Tambora 1815) before any negative result is trusted.
• Claude Code used collaboratively to build the workflow; the tiered System-One filter idea came from the engineer, not the AI.

📖 Read the full article
🔗 Source

#AIAutomation #AgentWorkflows #LLMOps #SemanticSearch #DocumentAI

Agent Standup: visualize Claude Code multi-agent sessions as a live standup call

Agent Standup turns Claude Code sessions into a real-time visualization styled as a team video call, showing each agent's status: working, waiting on another agent, needs approval, stalled, or done. It reads session files already written to disk, adding zero extra token cost and sending nothing externally. Free and open source.

• Per-agent live status (working, waiting, blocked, needs approval, done) sorted by urgency
• Reads Claude Code's existing session files — zero extra tokens, fully local, nothing sent anywhere
• Runs as a VS Code/Cursor extension, npx CLI, or desktop app on Mac, Windows, and Linux
• Extras: session replay with highlights, context gauge, video export, and optional voices

📖 Read the full article
🔗 Source

#AIAutomation #ClaudeCode #MultiAgent #AIAgents #LocalFirst

15 AI agents decompiling a game: practical lessons in orchestration, verification, and focus

A developer ran 14–16 AI agents (Claude Code + Codex CLI) for three months to decompile a first-person shooter to C++. The core finding: without a machine-checkable PASS/FAIL oracle, agents produced readable but semantically wrong code. Adding a byte-matching verification script fixed quality and made cheaper models reliable at scale.

• Agents tracked work via GitHub issues and communicated through a shared Discord channel, including CI failure webhooks.
• Early code looked correct but had wrong types, invented logic, and costly architectural changes — caused by no objective acceptance criteria.
• A byte-matching 'oracle' script replaced the human reviewer; strict automated feedback made cheaper models perform as well as expensive ones.
• Instructions decay: agents drifted and cut corners over time; an hourly cron job re-injecting the rules document kept them on track.

📖 Read the full article
🔗 Source

#AIAutomation #AgentOrchestration #LLMOps #ClaudeCode #AIAgents