Foundation models are important. So are memory systems, browsers, screens, policies, evaluation loops, and the humans who have to trust what happens next.
Long-term semantic recall, episodes, associative retrieval, consolidation, temporal context, user models, and world models. The research question is not simply “can it remember?” It is “can it remember the right thing, at the right time, with the right uncertainty?”
Accessibility-first perception backed by OCR and visual grounding, click memory, drift calibration, outcome classification, and recovery after interfaces change.
Playwright for structured operation, a real-Chrome bridge for real sessions, accessibility inspection, action caching, and page-drift recovery.
A meta-harness evaluates memory recall, tool selection, task completion, research quality, and token efficiency; explores candidate configurations; applies safety gates; and supports rollback. “Self-improving” only matters when the improvement survives evaluation.
ATLAS OS is the pressure test: company knowledge, monitors, playbooks, employee agents, approvals, audits, tenant boundaries, and real tools with real consequences.
We plan product-specific models, learned multi-model orchestration, and a general-purpose model intended for public use. Architecture, parameters, license, benchmarks, dates, and access are not announced.
Least privilege, human authority, reversibility, fail-closed boundaries, evaluation, isolation, privacy, auditability, prompt-injection resistance, and honest limits. Safety is not a separate lane. It crosses every one of the lanes above.
We are interested in useful problems, rigorous criticism, safety work, and collaborations that make agents more capable without making people less powerful.
Bring it to the lab →