Long Running AI Agents | On The Edge #4

Long-running AI agents are best understood as autonomous, stateful processes that keep working for hours or days—planning multi-step goals, calling external tools, and persisting progress via memory and checkpoints—without needing constant user prompts. The practical challenge is controlling how long the agent runs and ensuring it stops at the right moment, especially when tool calls and web research can keep expanding. This walkthrough focuses on a simple, time-boxed framework to make that control concrete.

The setup defines a “longrunning AI agent” as an autonomous stateful software process that can persist for a long window while maintaining memory and checkpoints. Instead of aiming for days, the demo uses short “time tasks” to prove the mechanism: the agent should research for a fixed duration, update a live web page in real time, and hard-stop when the time window ends. To do that, the system uses MCP servers as external tool providers—specifically custom MCP tools for searching X (including handles and popular posts) and for doing similar research on Reddit. Built-in tools like web search and fetch are also available, so MCP is used where it adds targeted capabilities.

A key operational rule is embedded directly into the prompt: for time-based tasks, the agent must adhere to the time spent and must never stop working before the time is up. The initial job is explicitly time-boxed to five minutes. The agent is instructed to perform five minutes of research on “Sora 2,” then produce an in-depth article with three paragraphs and bullet points, aiming to find a Sora 2 video to embed. Crucially, it must keep updating and refining the existing content rather than simply appending new text. A background timer (implemented with a sleep call) tracks the 300-second window, and the agent is expected to stop research exactly when the timer expires.

During the five-minute run, the live page updates as the agent performs web searches and MCP-backed queries against X and Reddit. When the timer hits “time is up,” the process stops as intended. The results show the agent generated content but failed to embed an actual Sora 2 video; instead, it produced a plausible-sounding but imperfect title and summary. The point, however, is less about perfect research and more about verifying that the agent respects the time boundary and can iteratively update a working artifact.

A second run extends the window to ten minutes using the same prompt structure, with the timer adjusted to 600 seconds. Again, the agent stops when the time expires, confirming the time-boxing behavior scales at least from five to ten minutes. The longer run adds more substance—an image appears—but also risks producing “too much information” relative to what the prompt asked for, suggesting that prompting and constraints still matter even when the runtime control is correct.

The demo closes by connecting this approach to broader claims about long-horizon agents. It references an Anthropic observation that Cloud 4.5 maintained focus for more than 30 hours on complex multi-step tasks, implying that longer durations are possible but require a stronger framework than a simple time-box. The takeaway is that time-based “longrunning” agents can be built with MCP tool calls, real-time artifact updates, and strict stop conditions—making it easier to experiment with longer autonomous workflows in a controlled way.