Building agents that plan, act and know when to stop.
AgenticWorks is a small independent lab exploring autonomous AI agents — the runtimes they live in, the tools they use and the evaluations that tell us whether they actually work.
What we work on
Practical problems that show up when language models stop chatting and start doing things.
Agent runtimes
Scheduling, memory and tool execution for long-running agents that survive restarts and failures.
Tooling & integrations
Safe, observable interfaces between agents and real systems — files, APIs, browsers and devices.
Evaluation
Task suites and metrics for multi-step behaviour: success rate, cost, recovery from errors.
Embodied & edge
Small robots and home devices as testbeds for agents that operate in the physical world.
Projects
Open experiments. Some are stable, most are moving fast.
AgenticOS
An experimental operating layer for agents: a process model, permissions and a shared memory space so several agents can cooperate on one machine without stepping on each other.
runtimepermissionsmemoryAI Lab
Our bench for trying new models and agent patterns: reproducible task environments, tracing and side-by-side comparisons of planning strategies.
evaluationtracingbenchmarksGet in touch
Collaboration, questions about the projects or just ideas — happy to talk.