Long-Running Agents

Long-Running Agents

Long-running agents are agent workflows designed to keep working across long horizons, pauses, approvals, external events, and restarts without losing state or confusing old conversation history for current workflow status.

Key points

  • Z.ai frames GLM-5.2 around coding and long-horizon tasks, including a 1M-token context window for extended work [src-095].
  • Google Developers frames the production problem differently: real agents need durable state machines, persistent session storage, event-driven waiting, and multi-agent delegation, not just a long context window [src-101].
  • The strongest pattern is to store workflow state explicitly. An agent should read its current step from session state rather than infer it from a long chat transcript [src-101].
  • For Robin's systems, this supports an operating principle: scheduled research, newsletter, and feedback agents should preserve explicit ledgers, checkpoints, and recovery paths instead of depending on one continuous prompt history [src-101].
  • Google for Developers adds the frontier view: a long-running model may be useful because of the quantity of independent work it can do, but that requires memory systems, continual learning, better hardware, low latency, and tools that are not designed only for human pacing [src-105].
  • The same source turns long-running agents into a full-stack problem: if the model is fast but browsers, APIs, and applications are slow, the agent spends much of its time waiting rather than working [src-105].

Related entities

Related concepts

Source references

  • [src-095] Z.ai – "GLM-5.2 | OpenLM.ai" (2026-06-13)
  • [src-101] Shubham Saboo and Eric Dong – "Build Long-running AI agents that pause, resume, and never lose context with ADK" (2026-05-12)
  • [src-105] Google for Developers – "Gemini co-leads on project origins and what's next" (2026-05-29)

Robin Cartier perspective

This page is part of Robin Cartier's working AI knowledge graph: a practical research layer for production AI, recommendation systems, experimentation, GEO, and agentic web readiness.

The useful next step is to connect this concept back to applied product leadership and operating models.

Recommended next

Keep reading from this thread

From 491 indexed pages and articles.

  1. Wiki concept Agent Tool Latency Bottleneck The pattern where a model becomes fast enough that the slow part of the workflow is no Related by running
  2. Wiki concept Google Agent Development Kit Or ADK, is Google's framework for building agents. In [src-101], the Google Developers Blog uses ADK to show Related by 101
  3. Insight AI Beyond POCs How enterprise AI moves beyond proofs of concept through ownership, governance, measurement, adoption, and production operating models Readers have engaged with this next