Opsle researches how much work in agentic AI actually requires model cognition and how much can be safely moved into deterministic systems.
Does an operation genuinely require cognition?
- Yes: give it to a model with the minimum sufficient evidence.
- No: implement it deterministically.
- Unsure: experiment, measure correctness and efficiency, then keep, change, or reject the mechanism.
Start with the research hub and the source foundation for the future Opsle website.
- Agent Trajectory Profiler
- Semantic Edit Protocol
- Durable Supervisor
- Event-Driven Agent Wakeup
- Context Firewall
- Decision Evidence Protocol
- Agent State Ledger
- Agent Scheduler Runtime
- Verifiable Agent Handoff
- Agent Routing Policy
- Agent Resource Claims
- Agent Discovery Control
- Agent Execution Authorization
- Controlled Agent Acceptance
- Agent Recovery Policy
- Ephemeral Agent Workers
Opsle Tasks is the future name of the integrated reference implementation currently under active development. Its migration into the Opsle organization is intentionally deferred. The independent research projects are designed to remain understandable, falsifiable, benchmarkable, and useful without depending on that product.
No result is called proven merely because a mechanism exists in an application. Experiments, negative results, failure-derived learning, and remaining uncertainty are part of the record.