Current Explorations
Forward-looking research threads where agentic systems meet operational constraints: memory lifetimes, orchestration failure, operator oversight, telemetry, and multi-agent coordination.
Systems under investigation
- EXP.MEMLab active
Agent Memory Systems
When to promote session scratchpad into durable memory, how to invalidate stale embeddings, and compaction policies that keep recall faithful without ballooning context.
- EXP.ORTLab active
AI Runtime Orchestration
Schedulers that respect dependency graphs across tools and models—cancellation propagation, partial replay, and backoff strategies when upstream latency spikes.
- EXP.HITLExploring
Human-in-the-Loop Architectures
Friction-aware escalation: surfacing confidence deltas, preserving operator intent in revision loops, and audit trails that survive compliance review.
- EXP.OBSExploring
AI Observability
Trace schemas that separate model latency from retrieval vs tool hops; budgets for tokens per outcome class; anomaly cues before user-visible regressions.
- EXP.MAGExploring
Multi-Agent Collaboration
Shared scratchspaces vs message-passing agents; conflict resolution when sub-agents disagree; role boundaries so autonomy doesn’t become chaos.