lsn-agent-v3 is a self-hosted agent for software engineering and IT operations. It proposes a plan, waits for your yes, works through hundreds of tools, flags its own unsupported claims — and runs on local models, on one machine you own.
Model and GPU figures refresh from the live system every 30 minutes.
Most AI agents rent you a mind in someone else's data centre. This one is yours: it runs where your systems run, sees what you allow, and never acts on its own beyond what you have explicitly made safe.
It was shaped on 2,326 real logged tasks and 37 runbooks from the lab it runs — the kind of work an infrastructure engineer actually does.
Health, security, backups and media checks run on a schedule; it reads the cluster through its monitoring tools and sends a short, sourced digest.
It pulls events, logs and metrics, correlates them, and proposes a fix as a plan — changes wait for your approval; reads never do.
Runs in parallel across the offline library of 123,000 files, offline Wikipedia and the live web, then writes a report whose citations are checked against the sources it opened.
A stored project advances one bounded step per trigger — across days if needed — and ends with a finalize pass that assembles the result before you see it.
Sketches of typical runs. The research figures are from a real run; the others illustrate the shape of the output.
The loop every task goes through. Up to 40 tool rounds per turn, streaming as it works, with every step and observation recorded.
A structured plan, validated against the tools that really exist — no imaginary commands.
You see the steps and their risk. Writes ask; critical actions also need a second factor.
Step by step, observations saved. When a step fails, it re-plans instead of guessing.
Evidence checks run on the answer; unsupported or invented claims are flagged before you read them.
At 70% of the model's real window, the conversation folds into a structured running summary — goal, decisions, facts, files, open items — and work continues.
A tool that returns megabytes doesn't drown the context: the output is saved as an artifact and only its head, tail and path stay in view.
Pinned notes join every session; search spans sessions, notes, projects and artifacts — diacritic-insensitive, so Romanian works as well as English.
When a task is bigger than one train of thought, the agent hands parts to fresh-context workers — each with its own persona, tools and bounded depth. Every sub-run is recorded.
28 project kinds — incident, fix, deploy, migration, audit, research, essay, translation, dossier and more — each define what "finished" means and which tool families apply.
A fresh worker takes a sub-task with a clean context and returns a result, not its whole transcript.
Independent questions fan out at once — the library, the encyclopedia and the web searched side by side.
Documents larger than any context window are split, read piece by piece and recombined into one answer.
Two mechanisms most agents don't have: deterministic evidence checks on every answer, and decisions that are read from the model's probabilities instead of its prose.
Illustration of the evidence footer, using the figures from the agent's real image audit. The flagged sentence is the kind of confident guess the checks exist to catch.
When the agent must choose — is this claim supported? is this action safe to run unattended? — the answer is read from the model's token probabilities over a fixed set of options. It can only pick one of the choices it was given; there is no free text to hallucinate. Every decision is audited, storing a hash of the situation rather than the text.
decide · "Is this action safe to run unattended?"
Illustration. Below the confidence bar, the safe choice wins by design.
An agent that can restart services and edit infrastructure must be predictable. The rules below are enforced by code and the operating system — not by asking the model nicely.
Logs, metrics, listings, searches. Runs freely — even when no human is around.
Asks first — once, for the session, always, or never: your call, per tool.
Recursive deletes, dropping data, force-pushes, deleting cluster objects: approval and a second factor, every time.
Built-in tools cover the engineer's bench; MCP servers connect it to the whole lab. A smart router ranks tools by meaning, so the model only ever sees the handful that fit the task.
files · shell · git · docker · kubernetes · helm · terraform · cloud CLI · network · ssh · PDF · DOCX · ODT · XLSX · CSV/JSON/jq · safe calculator · memory · projects · notify
Kubernetes · Prometheus · Loki · Grafana · network diagnostics · security · threat intelligence · the offline library of 123,000 files · offline Wikipedia · the Blender studio
Tools are ranked by embedding similarity to the request, nudged by intent hints and scoped by project kind. A search_tools meta-tool finds the rest when needed.
Repeatable procedures live as SKILL.md files in the open Agent Skills format — git-versioned and validated in CI: kubernetes triage, backup health, pod log scans, cluster self-heal, repairing its own tools, the studio. The agent can draft a new skill from experience; a human reviews it before it joins the library, which the agent itself can only read.
Ask for the network, a namespace or the whole cluster and it renders the diagram through its diagram tools — draw.io and Mermaid — grounded in what the cluster reports right now.
Behind the agent sits a sandboxed Blender with no network and 33 studio tools. Describe the thing; it builds it in millimetres, checks the numbers, looks at its own render with a vision model and hands back a file for the printer, the laser or the screen. All offline, on the same GPU that runs the models.
Parts from tested templates — stands, boxes with lids, hooks, brackets, coasters, tags — or shapes fused into one solid. STL and 3MF in millimetres with a printability report: one piece, nothing floating, wall thickness, overhangs, grams of filament. Repair and solidify when a mesh misbehaves.
SVG and DXF outlines with their holes, a section at any height, 16-bit heightmaps for relief carving. Every export reports size, shapes, holes and cut length, so the agent can tell you whether it fits the bed.
Realistic product shots on a wooden table, in studio, sunset or night light; turntables; a phone dropped onto the stand with real physics. Videos render as background jobs the agent polls — on the GPU, path-traced.
Realistic people from a CC0 base mesh, rigged on the fly; stunts that obey gravity, with slow motion at the apex; a night city; a title that gathers out of sparks; camera shots that follow the action. The first one — "a realistic person jumps over a building, with LSN appearing in a mesmerizing way" — was 300 frames at 1080p, rendered unattended in two hours.
Numbers first: pieces must be one, floating parts zero, the size what was asked. Then the render, read by the vision model — parts floating, cut off, inside a wall? It fixes them, and never says "done" while a check disagrees.
The studio's Blender runs in the browser, so you can watch it work from any device; a gallery shows every project; files land in a folder you control. The studio has no network at all — nothing leaves the machine.
The autonomous driver advances a stored project one bounded step at a time, across days if necessary, and closes with a finalize pass that assembles the report. Writes stay proposals. Here is what the first run produced.
:latest tags read 3 instead of 30, and a phase header said 15 items while listing 18. That is exactly why autonomous work ends as a report and proposals, not as changes.A test suite that quadrupled in ten days, research batteries run as an ordinary user, a full interface audit — and a candid comparison with a commercial model.
894 passing at the latest deploy · 224 commits since June 2026.
Every task completes with its file and a real answer in 1–25 minutes. Occasional factual slips remain — and the evidence footer flags many of them.
Zero console errors at desktop and phone widths, in dark and light themes; the rest fixed, partial or scheduled.
One AMD Radeon AI PRO R9700 with 32 GB, a dual-Xeon host with 88 threads and ~512 GB of RAM, and a llama.cpp router that wakes the right model in about 2.5 seconds from a RAM disk.
| Role | Model | Why |
|---|---|---|
| default · fast · tools | Qwen3.5-9B · MTP | quick, 8 parallel slots |
| heavy · long · decisions | Qwen3.6-27B | fully on GPU, 131k context |
| vision | Qwen3.6-35B-A3B · Qwen3-VL | screens, scans, diagrams |
| search & routing | bge-m3 | embeddings for tools & memory |
Every role can be pointed at another backend — local by default; external providers optional, per role, off by default. It is fully operational with local models only. The router serves 7 models today.
A single-page web app built for daily use, plus a command line, a scheduler, notifications and a REST API — all behind single sign-on.
It lives inside a Kubernetes cluster of 67 namespaces and 156 running pods — and it can see, explain and operate all of it.
It began as a personal assistant for one lab and grew, version by version, into an agent that treats safety and evidence as features — not afterthoughts.
Command line and chat, keyword routing between model sizes, a handful of personas and the first MCP servers. The seed of everything that followed.
A shared core with thin interfaces, plan-and-execute, a real database and tools that heal their own connections.
Local models on the new GPU router, multi-user sign-on, compaction, artifacts, sub-agents, project kinds and ~550 tools.
Typed decisions, risk levels with one-time codes for privileged work, evidence and citation footers, several audit rounds.
The SKILL.md library with reviewed promotion, the autonomous project driver and the narrow unattended lane.
The cluster image audit — completed unattended, reviewed by a human.
Open-weight models on your own hardware. Your logs, configs and documents never have to leave the building.
Risk levels, sandboxing, a separate privileged helper and out-of-band codes — enforced by code and the OS.
Gating choices are read from probabilities over fixed options. Nothing to hallucinate, everything audited.
Deterministic checks flag unsupported or invented claims right in the footer you read.
~550 tools across Kubernetes, observability, network, security, threat intelligence and an offline library.
Compaction, artifacts, sub-agents and resumable projects — work that spans days, not one chat window.
lsn-agent-v3 runs privately in the lab. If you lead an engineering or IT team and want sovereign AI that plans, asks and proves its work, I'd be glad to show you a live session.