Building model-independent agents for research & workflow automation. Sharing practical notes on agent graphs, tools, and evals. Follow for build experiments.
Joined October 2023
- Tweets946
- Following135
- Followers53
- Likes34
Pinned Tweet
Stop treating AI like a search engine. Start treating it like a team.
Chatbots answer a question and forget you exist the second you close the tab.
Agentlas actually remembers your endgame.
Give it one massive goal, and it:
• Breaks it down into daily sub-goals
• Assigns a squad of specialist agents to execute
• Saves progress at checkpoints so you never start from scratch
It doesn’t wait for your next prompt. It works toward your goal every single day.
agentlas.cloud
1
16
Introducing Agentlas 2.0.
Our biggest architectural update yet comes down to one premise: your agents should be real things you can see, own, and trust on your machine.
What changes in 2.0:
- One Home: One Team and Work merge into a single unified Home. Open 1:1 chats, group rooms, and coding sessions side-by-side in a 4-pane split workspace. Every group room now has a designated Lead agent, routing tasks transparently via @mentions with no hidden black-box coordinators.
- Native Sandboxing, Zero VMs: Downloaded agents live on your computer in their own local folders. Instructions, skills, and tools sit in plain sight. Each agent runs in an OS-level sandboxed terminal as a lightweight process that sleeps when idle and wakes on demand no heavy virtual machines required.
- Four-Tier Local Memory: To prevent context pollution, memory is physically separated into four local databases powered by local embeddings: Personal, Agent, Project, and Session. Your writing agent never sees your coding agent's notes, and side-project context never leaks into work.
- Self-Evolving Skills: When a workflow succeeds, your agent proposes it as a reusable skill. You approve it, and it joins the agent's folder.
- Cross-Device Continuity: One personal agent moves seamlessly across your computers, with your phone and the web acting as windows into it complete with mobile connectors for location, photos, markets, and news.
- Open Source: Licensed under Apache-2.0. Your data stays on your machine.
Coming soon to Agentlas Desktop. Read the first look:
1
2
62
Agentlas AI retweeted
The hard part of a big goal isn’t the work. It’s turning “grow my channel” into a plan you can actually check.
Agentlas OS does that for Claude Code and Codex.
Type a goal with a number and it becomes a Goal Tree:
the final goal
strategic goals, each with a KPI line
sub-goals, each with a “done when”
You can edit it, reorder it or pause a branch while it runs.
Then it finds an agent for each role from three places: your computer, your private Agent Cloud and a free public Hub.
Open source, Apache-2.0. One command to install.
If you find it useful, give it a star:
github.com/agentlas-ai/Agent…
#ad
Inspired by the recent wave of lightweight, zero-install virtual spaces, I built a browser-based 3D gallery.
Stripped away all the bloated metaverse extras no clunky avatars, no forced sign-ups. Just clean spatial exploration and art on mobile or desktop.
Check it out here: metaverse-gallery.pages.dev
Built on the web. Would love to hear your thoughts or feedback!
2
1
98
Agentlas AI retweeted
this week in open source ai: 17 new projects added, Sep 7-Sep 13, 2026 (Mon-Sun UTC).
1. qwen3.8-flash-next-in-c: Native C implementation of Qwen3.8-Flash-Next for local CPU inference, with terminal chat, a resident OpenAI-compatible API, and reproducible benchmarks.
github.com/shyringo/qwen3.8-…
2. Agentlas OS Agent Operation Environment (AOE): Local-first, model-agnostic environment for building, borrowing, and orchestrating specialist agents and teams across supported LLM hosts, with owner-scoped packages and host-enforced permissions. Apache 2.0 licensed.
github.com/agentlas-ai/Agent…
3. fractal: Hierarchical coding-agent orchestrator with recursive delegation, per-node Git worktrees, configurable limits, persistent SQLite state, and live terminal monitoring and steering.
github.com/plasma-ai/fractal
4. LWC (Local Wiki CLI): Local-first, source-grounded project memory for coding agents with bounded MCP retrieval, citations and provenance, atomic changesets, and optional document and code knowledge graphs. Apache 2.0 licensed.
github.com/JanYork/llm-wiki-…
5. Wenlan: Local source-backed AI knowledge base and LLM wiki for coding agents, with Markdown pages, hybrid retrieval, knowledge graphs, CLI workflows, and MCP integrations. Apache 2.0 licensed.
github.com/7xuanlu/wenlan
6. deja: Memory layer over the session transcripts coding agents already write to disk: Claude Code, Codex, Cursor, opencode, Zed and more, including sessions from before it was installed; local BM25 recall over them with no LLM or embeddings, MCP tools, and credentials redacted at index time. MIT licensed.
github.com/vshulcz/deja-vu
7. Busabase: Open-source database and workspace that gives AI agents structured data, durable knowledge, docs, skills, and apps through MCP, OpenAPI, CLI, and coding-agent skills; material writes can remain reviewable ChangeRequests before becoming canonical. MIT licensed.
github.com/busabase/busabase
8. Superagent: macOS desktop app giving Claude Code and Codex a real browser to drive, an iOS Simulator to install and screenshot apps in, and a phone companion app for remote monitoring. MIT licensed.
github.com/pungme/superagent…
9. Caura: Open-source memory infrastructure for AI agents, providing persistent memory and context management for building stateful, long-running agents.
github.com/caura-ai/caura
10. Unlimited Context: Local-first context and memory engine that keeps a billion-token encoded pool on disk and pages relevant slices into a small resident window for Ollama and other local LLMs. Apache 2.0 licensed.
github.com/AetherAI3/Unlimit…
11. inspeximus: Memory layer that retires a corrected fact by key and can undo the correction later from the key alone. Zero-dependency core with an MCP server.
github.com/DanceNitra/inspex…
12. Mnemoverse: MIT-licensed MCP server for agent memory, backed by a hosted engine that re-ranks recall from reported outcomes rather than similarity alone, with one key shared across Claude Code, Cursor, VS Code, and ChatGPT.
github.com/mnemoverse/mcp-me…
13. OrcaReplay: Records a coding-agent run below the harness and replays it offline with the network off, or forks it from a checkpoint onto another model.
github.com/Continuum-AI-Corp…
14. YYLO Benchmark: Longitudinal evaluation for agent runs: executes task prompts in private fresh-repository attempt workspaces and retains attempt evidence, evaluator provenance, and reports for later re-evaluation. MIT licensed.
github.com/yylo-dev/yylo-ben…
15. AI Segmentation by TerraLab: QGIS plugin for point-and-click SAM segmentation of buildings, trees and any object in satellite and drone imagery, with a free CPU-only local mode.
github.com/TerraLabAI/QGIS_A…
16. SpecJudge: CLI that reads a project's specification files and recommends which LLM fits the work, using a local judge model via Ollama and citing the spec fragment behind each demand level. MIT licensed.
github.com/JoaquinRuiz/SpecJ…
17. YYLO: Command-line orchestrator that runs AI coding agents (Claude Code, Codex, Gemini CLI) in parallel across isolated git worktrees, tracked on a Kanban board with typed merge and release flows. MIT licensed.
github.com/yylo-dev/yylo
explore the full open-source ai list at awesomeosai.com/
Agentlas AI retweeted
A multi-agent workflow needs more than a pile of prompts.
Agentlas OS is an open-source agent OS for builders who use agent hosts such as Claude Code, Codex, Cursor, and Gemini.
It helps you turn a request into a staffed workflow by splitting it into tasks, bringing back specialist candidates, and keeping planning, execution, synthesis, and verification separate.
Key features:
• Task splitting – turns a sentence into roles, handoff order, and an expected artifact for each task
• Candidate sourcing – brings candidates from your machine, private cloud, and the public Hub for each task
• Qualification controls – supports declared, checked, demonstrated, and attested qualification levels with task-specific floors
• Verification receipts – requires an independent verifier and records where a run stopped when work cannot be completed
• Host integrations – provides runtime adapters for Claude Code, Codex, Cursor, Gemini, and other supported hosts
It’s open-source (Apache License 2.0).
Link in the reply 👇
We do NOT have any official token or coin. Any account claiming $AGENTLAS with a contract address is a scam.
2
27
Agentlas AI retweeted
Most people are still using AI agents like chatbots.
That's why they're missing the bigger picture.
I tested @MasonL377167, and it's a completely different approach.
I gave it one objective:
Build an AI automation agency in 60 days and get 10 paying clients.
It turned that goal into:
→ Strategic goals with measurable KPIs
→ Actionable sub-goals with completion conditions
→ A structured 60-day execution roadmap
But that's just the beginning.
I also explored its Agent Hub, which discovered specialist agents for:
→ Market research
→ Lead generation
→ Sales strategy
→ Growth marketing
You can find agents, build your own, and coordinate multiple specialists around a bigger objective.
And the best part?
It's open-source and works with Codex, Claude Code, Cursor and Antigravity.
If you're a developer, this is worth exploring.
Install it, experiment with it, and see what you can build.
GitHub: github.com/agentlas-ai/Agent…
Star the repo if you find it useful.
#Ad #AIAgents #OpenSource
🤝 Paid partnership
Come to an online yoga studio and practice singing bowls and mind training together.
हमारे ऑनलाइन योग स्टूडियो से जुड़ें और सिंगिंग बाउल की ध्वनि के साथ मन की साधना साथ मिलकर करें।
Mari sertai studio yoga dalam talian kami dan amalkan ketenangan minda bersama alunan singing bowl.
Hãy tham gia phòng yoga trực tuyến để cùng lắng nghe âm thanh chuông xoay và rèn luyện tâm trí nhé.
hush.hush-realtime.workers.d…
3
38
Just shipped Local Models support in Agentlas 🚀
No more guessing which model fits your machine. Agentlas detects your hardware specs (like Apple Silicon unified memory) and surfaces the best-fit GGUF models from Hugging Face for seamless one-click downloads.
Making local AI agents truly plug-and-play.
#Agentlas #LocalLLM #OpenSource #AI #BuildInPublic
2
3
119
Agentlas AI retweeted
What if your AI could turn a conversation into a reusable agent and find other agents whenever it needs them?
Meet Agentlas OS, an open-source agent system that brings agent building and automatic discovery to your AI workflow.
🔹 Build Session Turn a working AI conversation into a reusable agent with /hep-build session.
🔹 HEP Network Automatically discover the right agents for each task across local, private cloud, and a public agent network.
🔹 No more agent catalogs Stop manually installing individual agents or feeding your AI endless lists of skills.
Build an agent once. Save it. Share it. Let other AI agents find and use it when needed.
Works with Claude Code, Codex, Cursor, and other AI environments.
Agentlas AI retweeted
2023 AI: "Here is a 500-word essay on how to fix your Excel errors."
2026 AI: *Literally grabs your mouse and finishes your entire workday in few minitues.*
Spreadsheet cleanup → 5-slide deck → Team chat update → PSD background removal → Video edit → 3D render.
6 prompts. Handled entirely by SI Agent inside Agentlas One.
Agentlas AI retweeted
Most AI coding agents wait for you to open a file.
We built an agent that maps your entire repo in 3D and fixes the files you forgot existed.
1. Lay out the canonical folder & file tree across a 2D (XY) grid.
2. Extrude the Z-axis using Git touch frequency (folder + file access counts).
3. Deploy the agent starting from the lowest Z-coordinates—auditing, refactoring, and fixing the coldest code first.
By turning Git history + directory structure into a 3D priority landscape, the agent systematically eliminates silent bugs and dead weight from the bottom up.
Meet the autonomous codebase janitor inside Agentlas. 👇
Most AI coding agents wait for you to open a file.
We built an agent that maps your entire repo in 3D and fixes the files you forgot existed.
1. Lay out the canonical folder & file tree across a 2D (XY) grid.
2. Extrude the Z-axis using Git touch frequency (folder + file access counts).
3. Deploy the agent starting from the lowest Z-coordinates—auditing, refactoring, and fixing the coldest code first.
By turning Git history + directory structure into a 3D priority landscape, the agent systematically eliminates silent bugs and dead weight from the bottom up.
Meet the autonomous codebase janitor inside Agentlas. 👇