All your agents are going async

· ai programming · Source ↗

Article

TL;DR

Agent tasks outlive HTTP sessions — durable transport and session state are now required primitives.

Key Takeaways

  • HTTP request/response breaks when agents run hours, switch devices, or fire unprompted
  • SSE + Last-Event-ID with durable state may solve reconnection without heavy new infrastructure
  • Author discloses conflict: post reads as pitch for Ably; community skeptical of the specific solution

Discussion

Top comments:

  • [bozdemir]: Transport isn’t the problem; interaction model is — most async use cases just need a Slack message

    The transport isn’t really the problem though. The interaction model is. When my cron agent finishes a task at 3am I don’t want a live session I can rejoin, I want it to drop a message in Slack.

  • [TacticalCoder]: Simplest durable channel for async agents is $5 VPS, SSH, and tmux
  • [pdp]: Injecting messages into agent context mid-run doesn’t reliably trigger action — architecture flaw
  • [tuo-lei]: Real blocker is non-portable session state between tools, not transport layer

Discuss on HN