01 / 03
clip coming soon
Non-stop manager
The agent works while you're away; decisions wait in a tray, not in the chat.
/ Przemysław Tarkowski
I built a production system of autonomous AI agents — and I run it every day.
Senior engineer, 11 years shipping real software — a 100M+ download game, solo-built AWS infrastructure. Now aimed at AI agents, with a working system to show, not slides.
TarkOS · open source — coming soon
layer
Claude Code harness
agent OS
audito (agent OS)
01
Decisions (Decision-OS)
02
Memory / capture
03
Non-stop autonomy
- tools built
- 140+ tools built
- skills
- 50+ skills
- sessions run
- 280+ sessions run
- every change
- eval-gated every change
/ What the system does
Not a chatbot — an operating system for getting real work done autonomously.
- 01 autonomous orchestration
Runs on its own
Give it a goal and it works across sessions while I'm away — inventing the next task, not waiting for prompts.
- 02 human-in-the-loop · governance
Decides like me — asks only when it must
It owns the reversible calls and batches the irreversible ones into a decision tray for my sign-off.
- 03 eval-driven
Measures its own work
Every change is regression-gated and retrieval is scored — the agent grades itself before I ever see it.
- 04 self-evolving
Gets better every session
It captures my corrections into explicit rules, so the same steer is never needed twice.
Note The code lives inside my private work harness (open-sourcing soon). I'm happy to walk you through it live — architecture, decisions, the eval loop.
/ Case study · the system in depth
How I run agents unattended — and still trust the output
The problem
An agent left to work alone breaks the same way every time: it makes decisions it can't justify, does irreversible things without asking, and loses the thread the moment a context window fills. The gap between an impressive demo and a system you'd actually leave running is engineering, not prompting.
- Act 1 multi-agent · long-horizon
Orchestration & continuity
One orchestrator spawns and supervises worker agents across parallel workstreams, monitoring by work-signals rather than guesswork. When a worker approaches its context ceiling it relay-hops — closing cleanly and handing a successor the full thread (warm resume + lossless history load) — so long-running work survives across sessions, attended in the day or unattended overnight.
- Act 2 human-in-the-loop · governance
Decision-OS — the trust layer
Every decision is classified: reversible ones the agent makes itself, irreversible ones it batches into a tray for my sign-off (default-deny). An append-only, gap-detecting audit trail records each call with its rationale — tamper-evident, the way a regulated environment needs it. This is what turns 'runs by itself' into 'runs by itself, safely'.
The result — operated, not demoed
- sessions operated
- 530+ sessions operated
- fully-unattended runs
- dozens fully-unattended runs
- decisions logged w/ audit trail
- 120+ decisions logged w/ audit trail
- tools built around the loop
- 143 tools built around the loop
/ Where this comes from
Eleven years shipping — now pointed at agents.
The EU research I worked on was literally about autonomous orchestration of cloud/edge resources. Today I orchestrate autonomous agents. Same instinct, new substrate.
- 11 yrs
- shipping production software as a senior engineer
- 100M+
- downloads — Crazy Kick, published with Voodoo (LiveOps at scale)
- 2019–22
- solo-designed & ran AWS multiplayer infrastructure (EC2/ECS/Lambda/DynamoDB)
- EU H2020
- autonomous orchestration of cloud/edge resources (ACCORDION · CHARITY)
02 / Capabilities
A short clip per feature: what it does and how it works.
01 / 03 
02 / 03
clip coming soon
Decision queue
A queue of decisions with consequences and a recommendation. One click instead of meta-work.
03 / 03
clip coming soon
Capture / memory
Knowledge from every session lands in the right file with provenance. It's git-blame for knowledge.
03 / Skills layer
Agent systems, end to end
06 itemsWhere I'm strong today: the operating layer most agent projects lack.
- Agent architecture: rules, skills, knowledge, memory
- Decision routing with human-in-the-loop gates
- Autonomous lifecycle: overnight runs, watchdogs, recovery
- Multi-session orchestration on parallel workstreams
- Verification by outcomes: tests as contracts, regression suites
- Memory and capture with measured retrieval quality
Market baseline
04 itemsWhat I'm building on top of that. Today AI writes the implementation under my architecture; I verify by outcomes.
- Python AI-directed today
- RAG / vector DB own evals built
- LangGraph learning
- React soon
05 / Connect
Open to interesting conversations: AI engineering, agent systems, devtools, applied AI.
If you're curious how all of this works together, or you want to talk about agents, devtools, or applied AI, get in touch. Happy to walk you through it.