Verified Agent Skill record

Latent Briefing

Latent Briefing shares memory at the representation level by compacting the orchestrator trajectory directly in the worker model's KV cache, retaining positions most relevant to the current worker task rather than replaying the full trajectory as text or using lossy summarization.

Tasks

Define the shared memory boundary by deciding exactly what enters the orchestrator trajectory cache (prior worker replies, tool output, chain-of-thought, or selected artifacts)
Tune compaction aggressiveness as a controlled parameter on validation data, tracking task accuracy, worker tokens, total tokens, retention rate, and compaction overhead together

Inputs

Worker task prompts (used to derive task-guided query vectors)
Orchestrator reasoning trajectory
Long-document QA workloads and validation data

Outputs

A compacted KV cache for the worker model retaining positions relevant to the current task
Metrics covering worker-token reduction, total-token savings, compaction overhead, and task accuracy

Limitations and checks

Not viable for API-only stacks where internal KV tensors are inaccessible
Direct representation sharing may not work cleanly across unrelated model families with different tokenizers, architectures, or attention layouts
One global compaction threshold rarely works across long vs short context and easy vs hard tasks; accuracy cliffs can occur when compaction is too aggressive
Benchmark evidence focuses on long-document QA; code generation, math, and multi-document synthesis may behave differently
Confirm the runtime can inspect and rewrite worker KV state; if not, this skill cannot be implemented directly
Verify orchestrator and worker share a compatible model space (tokenizer, architecture, attention layout) before attempting latent handoff
Compare end-to-end latency of compaction plus generation against strong text-level baselines such as prefix caching, structured notes, retrieval, and selective text handoff
Average results over enough trials to separate compaction effects from orchestrator decomposition variance