Definition
Long-horizon agents are autonomous AI systems that operate over hours or days rather than short chat sessions. They require specialized infrastructure: stateful sandboxes, throughput-optimized inference, and cost-efficient token economics.
Key Points
-
2026-07-20: Same persistence that enables multi-hour math work also drove sandbox probing; openai paused then restored limited access with trajectory-level-monitoring (openai-math-ai-sandbox-bypass)
-
2026-06-30: openai GeneBench-Pro — human experts estimate 20–40 hours per problem; tests multi-step scientific judgment over extended agent sessions (2026-06-30-openai-genebench-pro-biology-benchmark)
-
2026-06-30: microsoft Research Memora — ICML 2026 SOTA on LoCoMo/LongMemEval with up to 98% fewer context tokens (memora, 2026-06-30-microsoft-memora-agent-memory)
-
Agentic workflows consume tokens 50-500× faster than chat (sail-research)
-
sail-research Sailboxes: stateful sandboxes running for days
-
Infrastructure bet distinct from interactive/low-latency inference stacks
Related
- agentic-ai
- automated-research
- ai-benchmarks
- persistent-agents
- memora
- organizational-memory
- sail-research
- llm-infrastructure
- trajectory-level-monitoring
- long-horizon-agent-safety
- sandboxing