missionSTACK
Shared context, coordinated fleets and bounded work-cell actions. Explore missionSTACK in missionSTUDIO, our early-alpha workspace.
Explore →Loading…
Systown · Stockholm
Intelligence needs control.
We build the framework that connects AI to accountable action. Declared states. Bounded decisions. Visible evidence. From agent workflows to fleets, arms and grippers.
What we build
Two products. One principle: control.
missionSTACK for physical operations. contextCLONE for governed agents. Test before acting. Keep people in control.
contextCLONE
Approved sources. Permitted tools. Checked transitions. Agent workflows designed for human review and versioned improvement.
Explore →Models with a declared scope
Inspect the tested artifact, fallback and limits. Generation-3 naming does not turn simulation evidence into hardware performance.
Explore →
Our approach / State machines & agentic loops
Declared states. Bounded decisions.
Models propose. Rules admit the action. The framework controls the loop.
Declare the loop
Name states, actions and stop rules. Bound rule, reflex, model and human decisions. LangGraph orchestrates missionROS research; the plain controller owns the control tick.
Explore →Keep the record
Record sources, checks, decisions and overrides. Replay the same version and recorded outputs without new model calls.
Explore →Improve through a gate
Replay failures. Evaluate candidates. Review the evidence. Approval creates a new version; the running policy stays under human authority.
Explore →
swarmREFLEX3 / swarmREFLEX3-ROS
Fast responses. Visible evidence.
Handle routine work locally. Reason about exceptions. Keep the experiment behind every claim.
Fewer world-model calls
98.3% median reduction across 16 paired simulation seeds, from coordination plus rule reflexes. Supports the up-to-90% headline; not a learned-model-only result.
Explore →Local decision timing
Earlier model-v4 path: 24.75 ms p99 on DGX Spark CPU. Separate Jev/Laya tests disclose admissions and deployment paths.
Explore →To arms and grippers
Six contracts bound work-cell responses. Learned hold research is kinematic; hardware timing and commissioning require separate evidence.
Explore →
People
Accountable for the work.
Patrik Johansson
Chief Executive Officer
Pontus Sahlberg
Chief Technology Officer
Annika Johansson
Chief Operating Officer
Magdalena Brolin
Head of R&D and Development
Kristian Pettersson
Edge AI System Lead
Ella Tojit
Physical AI Platform Lead
Oscar Hansson
Agentic Physical AI Lead
Edwin Andermyr
Agentic Platform Lead
Rebecca Jacoby
Physical AI Use Case & Test Engineer
Krister Folkeson
Chief Financial Officer
How we work
Build boldly. Prove carefully.
Build at the frontier
Build and evaluate with OpenAI on approved material. Keep model choice and data boundaries explicit.
Ground the physical world
Evaluate NVIDIA compute, AVEVA PI/AF history and Wittra spatial context against the site’s permissions, calibration and controls.
Prove before operating
Show assumptions, results and approvals. Verify sovereign and air-gapped profiles on the installed system.
Explore →
Start with your operation
Build what comes next.
Bring the mission. Define the controls. Build the evidence.