←── back to feed
/topics/long-horizon-agent-frameworks-and-evaluation

Long-horizon agent frameworks and evaluation

4 items2 sourcesupdated 24d agotrend 0

Three new arXiv papers address reliability and scalability challenges in long-horizon autonomous agent systems: Agent Mesh documents production failures in orchestrator retry logic, SKILL.state proposes a stateful runtime to replace append-only conversation histories, and PILOT introduces live self-improvement that redirects active runs rather than waiting for post-execution analysis.

  • Agent Mesh analyzed 147 production incidents across 81 runs, finding all three service-mesh assumptions (retry, timeout, circuit breaking) violated in agentic systems
  • SKILL.state replaces growing conversation history with mutable execution state to prevent latency degradation and context-poisoning failures over long horizons
  • PILOT enables live self-improvement during execution to redirect active runs and validate lessons immediately, rather than only post-execution processing
  • Group size effects and collective misalignment in LLM multi-agent systems published in PNAS on same date