Bartłomiej Szczygło · AI Engineer & Solo Founder · Warsaw
I work on the parts of agent systems that break the moment they leave the demo: giving an agent a browser, letting it use real accounts without handing it the passwords, surviving prompt injection, keeping inference up when a provider fails, and workspaces that persist between runs. Then I write down what broke.
Currently building
A runtime for AI agents: persistent desks, a credential vault agents can use but never read, and a browser, shell and API they drive themselves — with guardrails that fail closed when the content they read turns hostile.
What it is and why it exists
Most multi-agent systems fail because the control flow itself is generated by a language model. Invert it: a deterministic orchestrator that only routes, retries and times out, with all judgement pushed to the leaves where it can be observed and replaced.
A stateless agent re-clones the repo, re-installs the dependencies and re-derives the same conclusions on every run. Giving it a durable workspace — a desk that survives the session — removes more failure modes than a bigger context window ever will.
An agent that uses your GitHub, inbox and cloud console does not need to know a single credential. Put a broker between the agent and the account: the agent states intent, the broker holds the secret, and the key never enters the model's context.
The runtime an autonomous agent actually needs: persistent desks, browsers, shells, orchestration, recovery.
Context windows, tool design, structured output and streaming UI — the interface between a model and real work.
Routing, fallback engines, local models, cost control, evaluation and the boring plumbing that keeps inference up.
Prompt injection, fail-closed guardrails, untrusted content and keeping secrets out of a model's reach.
Adoption, process design and ROI — what happens when agents touch a real company instead of a benchmark.