Ask HN: Multi-agent workflows in production; Where people using 1000s of agents?

I feel like most AI workflows can be solved pretty effectively by a single capable LLM or with upto 5 subagents however, many engineering teams are focused on multi-agent architectures at huge scales.

Curious to understand exactly when it becomes worth it / what production use cases there are for large multi-agent swarms: I’m trying to understand exactly where that value lies.

If you are run agent swarms in production: what is the main use case / need and what is your biggest pain point right now (state sync, token costs, cascading failures, latency)?

[As context: I am a founder at Acyclic Labs (YC F26) and we are building infra to scale agents. Looking to map out when the swarms are actually justified and when they are wasteful!]

2 points | by ramstar3000 1 day ago

6 comments

  • taurath 20 hours ago
    I’d love examples of it actually working but right now all it’s seemed to be is hype.
    [-]
    • ramstar3000 12 hours ago
      any interesting hype you have seen at least lol
      [-]
      • taurath 7 hours ago
        I haven't found anyone who isn't working for OpenAI/Anthropic, or isn't selling a solution. There seems to be very little organic takeup - either out of budgetary reasons, or more likely ineffectiveness relative to ephemeral jobs.
  • kaihwang 1 day ago
    most of the pain at scale isn't the agents themselves, it's observability. once you're past a handful you basically need per-agent tracing or you're debugging blind, worth borrowing tooling from how people do this for microservices
    [-]
    • Var1377 22 hours ago
      does anyone have any recommendations for stronger multi-agent observability. I've not really seen anything outstanding in this space yet
      [-]
      • idempotent_ 22 hours ago
        Depends on your setup but we've had success with building custom harnesses and processing metrics via OpenTelemetry + Prometheus. This is running on a massive K8s cluster spinning up pods per N agents.

        If you are using a proprietary harness like CC or Codex then look into automatically ingesting their memory logs and internal datastores to track progress and then instrument the host to send back metrics

  • syndred 3 hours ago
    [flagged]
  • paul_irolla 12 hours ago
    [dead]
  • syncday 1 day ago
    [flagged]
  • vancekai 9 hours ago
    [dead]