How COSMA works
Keep the useful work from falling apart.
COSMA is an independently developed operating architecture for work that has to survive changing tools, sessions, drafts, decisions, and handoffs. The goal is simple: keep the source, the current state, the boundaries, and the next valid move connected.
A useful difference
The AI is not the authority.
Models can generate, analyze, compare, and execute. COSMA keeps those capabilities underneath a separate layer that tracks what actually governs the work.
| What matters | How COSMA handles it | Why that helps |
|---|---|---|
| Your source stays primary | COSMA keeps the current source, approved direction, and project state tied to the work instead of treating every new chat or tool as a fresh start. | Less re-explaining. Less accidental drift. |
| Changes stay traceable | New work continues from a known prior state. Rejected or failed directions can be preserved without quietly becoming the new baseline. | You can recover the work without guessing what changed. |
| AI output is not automatically approval | Generated material remains a candidate until the human decision-maker accepts it for the role it is meant to play. | You keep control of what becomes real, current, or final. |
| Different tools can serve the same project | Models and production tools are treated as processors, not as the owner of the project or its memory. | The project can survive tool changes without losing its identity. |
| Mistakes become useful evidence | A failed route is isolated, inspected, and corrected from the last valid state rather than silently reused. | The same preventable failure has less room to keep coming back. |
Try the lens
What keeps breaking?
Choose the closest pattern. This is not a diagnostic or sales quiz; it simply shows how COSMA frames common continuity problems.
Continuity problem
What is happening
The useful work exists, but the next session or tool cannot reliably tell what is current, approved, or still open.
What COSMA changes
COSMA carries a bounded current state: the active object, the valid prior state, the source that governs it, and the next move.
Useful outcome
The next session begins from earned ground instead of reconstruction.
Evidence snapshot
Built, tested, and still being tested.
COSMA keeps internal evidence separate from marketing language. The public view below reports implementation and test status in plain terms.
Versioned operating kernel
Implemented
Kernel records preserve prior versions rather than overwriting history.
Agent control plane
Tested
37 automated checks cover authority, signed production gating, rollback, tamper detection, and role separation.
Operational intelligence compiler
Tested
Deterministic current-state compiler with 20 executable self-tests and a larger adversarial acceptance specification.
Project continuity
Active use
Project capsules, causal-parent recovery, evidence-only rejected branches, and bounded writeback are used across live COSMA work.
External research comparison
Active research
COSMA is compared against current public work in long-horizon agents, memory, rollback, safety, and human control.
Research context
We compare the architecture, not the hype.
COSMA's internal methods are compared against current public work in long-horizon agents, memory systems, rollback, safety, and human control. These sources are references for comparison; they do not author COSMA.
Microsoft Research
Memory as execution-state management for long-horizon agents
Microsoft Research
CORPGEN: hierarchical planning, memory isolation, experiential learning
OpenAI
Safety and alignment in an era of long-horizon models
Anthropic
Effective harnesses for long-running agents
Google DeepMind
Gemini Robotics 2: whole-body intelligence, uncertainty, and human intervention