====== Long-Horizon Agents ======
Long-horizon agents are autonomous systems designed to execute extended multi-step tasks spanning hundreds or thousands of actions while maintaining coherent goal pursuit.(([[https://arxiv.org/abs/2509.24230|ELHPlan: Action Chains for Long-Horizon Planning]])) These agents must cope with context drift, compounding errors, sparse feedback signals, and the need to recover gracefully from failures deep into execution.

===== Overview =====
Most current LLM-based agents work well for short tasks of 5-20 steps. However, real-world problems — software engineering projects, scientific experiments, complex data analysis — often require sustained effort over hundreds of steps. Long-horizon agents address the fundamental challenge of maintaining task coherence and making progress over extended execution periods where the probability of encountering errors approaches certainty.(([[https://www.emergentmind.com/topics/long-horizon-agent-planning|Emergent Mind: Long-Horizon Agent Planning]]))

The key insight driving recent research is that naive flat planning fails at scale. Instead, long-horizon agents require hierarchical decomposition, persistent memory, proactive validation, and robust error recovery mechanisms. Contemporary approaches emphasize a paradigm shift from "one-turn cleverness" to sustained, iterative capability measured over hours of continuous operation.(([[https://thesequence.substack.com/p/the-sequence-radar-841-three-model|The Sequence - Radar: Long-Horizon Execution]]))

This transition reflects a fundamental change in how AI systems are evaluated and deployed. Traditional benchmarks assess single-prompt performance — writing a paragraph or solving a code snippet in one turn — but economically useful work typically requires planning, testing, and iterating over extended periods. Long-horizon execution represents sustained endurance and iterative problem-solving that better mirrors real-world demands, moving beyond isolated instances of "one-turn cleverness" to demonstrate reliable capability across hours of continuous operation. Recent systems like [[kimi_k2_6|Kimi K2.6]] have demonstrated this capability through extended autonomous infra agent runs and kernel rewrites, handling 4,000+ tool calls and 12+ hour continuous runs.(([[https://news.smol.ai/issues/26-04-20-not-much/|AI News (smol.ai) - 2026]]))

Success in this domain increasingly depends on massive context windows, sophisticated tool usage, and the ability to maintain state and coherence across complex, messy real-world engineering problems. The most advanced long-horizon systems require robust memory, runtime, and orchestration infrastructure to support capabilities like 300 parallel sub-agents operating concurrently.(([[https://www.latent.space/p/ainews-moonshot-kimi-k26-the-worlds|Latent Space - Moonshot: Kimi K2.6 (2026]])) Models emerging specifically for long-horizon execution prioritize endurance and iterative problem-solving.

===== Persistent State Management =====
A critical innovation in long-horizon agent systems is the approach to maintaining state across extended sessions. Traditional agent architectures lose context between interactions, requiring developers to implement explicit state serialization and retrieval mechanisms. Advanced persistent agent architectures abstract this complexity by providing virtual computing environments that function as persistent workspaces. Agents can leave work in progress, with the system preserving file states, application login sessions, and execution context for resumption.(([[https://www.theneurondaily.com/p/google-ran-out-of-cloud|The Neuron (2026]]))

This persistent approach enables several advanced capabilities:

* **Session continuity**: Agents maintain authenticated sessions across 1000+ applications without re-login protocols
* **Autonomous handoff**: The system manages transitions between scheduled execution windows, with agents automatically resuming work where previous sessions concluded
* **Stateful reasoning**: Agents can build operational understanding across multiple work sessions, accumulating context about task progress and application states

===== Practical Applications =====
Long-horizon agents target workflows that require continuous monitoring and autonomous execution. Common use cases include:

**Inbox Management**: Agents autonomously process incoming messages, triage communications, execute appropriate responses, and maintain inbox state across multiple review sessions.

**Travel and Booking Automation**: Multi-step booking tasks spanning multiple sessions and platforms, requiring persistent state maintenance across partial completions and external API interactions.

===== See Also =====
  * [[agentic_ai|Agentic AI]]
  * [[long_horizon_rl|Long-Horizon RL for Agents]]
  * [[ai_agents|AI Agents]]
  * [[agentic_workflows|Agentic Workflows]]

===== References =====