Evaluation and review controls for ambiguous agent work
Agent evaluations can test reliability more precisely by locating clarification, formal verification, and specialist review at the decisions where errors become expensive to reverse.
Ideas
Robot learning changes at explicit execution interfaces
Robot deployment teams can make imagined practice and real-world recovery data more useful by preserving explicit decisions that can be executed, checked, and corrected.
Executable controls for code optimization and generated tests
Performance and test-generation workflows can make model output more dependable by combining complementary executable signals: runtime profiles to prioritize static optimization matches, semantic mutations to challenge…
Counterfactual and trace-based tests for robot world models
Replayable episode twins can supply counterfactual supervision that recorded robot videos lack, while spatial traces can turn whole-episode replay failures into repairable alignment, contact, and dynamics errors.
Repository-aware editing, recovery, and evaluation for coding agents
Coding-agent controls can move closer to the semantics of the work: requirement links can constrain cross-file edits, invariant violations can improve recovery decisions, and controlled code transformations can reveal…
State representations and checks for embodied control
Embodied-control teams should preserve different information at different rates: dense spatial detail for the current scene, compact physical records across time, and independently refreshed state for execution checks.
Verification-aware cleanup of coding-agent work
Coding-agent cleanup should preserve the evidence needed to merge a change, not merely its passing status. The most useful changes are to make patch minimization coverage-aware, protect explicit obligations during…
Training changes for reliable long-horizon VLA execution
Longer memory and predictive representations need supervision tied to task meaning, physical constraints, and smooth execution.
Coding-agent controls tied to verification failures and harness changes
Repository exploration can be deferred until verification identifies a concrete knowledge gap, reducing unnecessary context while preserving a path to deeper repair.
Privacy controls for persistent, tool-using agent workflows
Pre-model PII handling does not cover every place an agent can retain or act on sensitive data. Repository traces, persistent memory, and connected applications need separate checks, while transformed identities should…
Capacity-aware controls for agent workflow reliability
Agent workflow operators can treat usage capacity and metering faults as execution conditions rather than external service incidents.
Closed-loop VLA training and evaluation changes
Phase labels can govern both sensor access during manipulation and compute reuse during policy evaluation. The evidence supports narrower changes to training and test infrastructure rather than claims of general…