Most agent demos optimize the happy path: generate a plan, call tools, return a result. Long-running work exposes a harder contract.
A useful system needs:
- source-linked memory rather than an unbounded prompt;
- freshness and correction paths for changing facts;
- idempotent execution receipts;
- explicit stop rules for new access, material judgment, and external commitments.
This separates preparing the next step from silently assuming authority to complete it. Proactive Execution becomes safer when the boundary is represented in the workflow, not left to a vague system prompt.
Klik is an app-first, pre-launch exploration of that architecture: Living Memory that keeps learning the working context, followed by permission-aware execution that moves routine work forward.
Top comments (0)