Skip navigation
What breaks your AI agents?
0:00 / 1:01

What breaks your AI agents?

GREG ISENBERG

@gregisenberg709K followers

Visit on X
112 likesDownloadOriginal post

29,231 views

What breaks your AI agents?

33 Comments

Comments from the original post on X (Twitter)

  • Notion's profile picture

    @NotionHQ

    hi friends @Shopify @github @figma @whatagraph

    10 likes

  • GREG ISENBERG's profile picture

    @gregisenberg

    @Shopify @github @figma @whatagraph we are a happy family

    2 likes

  • Nagajyothi's profile picture

    @njpCoder

    Love the video

    1 like

  • 泰坦链 | DeFi's profile picture

    @TiTan_Web3_

    The model gets blamed; stale state and oversized permissions usually did the damage

    1 like

  • Hussain Hashim | Building SundayBack's profile picture

    @itsthedonhashim

    @gregisenberg trying to get them to handle unexpected inputs without breaking. learned the hard way that flexibility is key.

    1 like

  • Adrian Karvinen's profile picture

    @adriankxyz

    Another banging video

    1 like

  • Promptyx's profile picture

    @promptyx_ai

    I think my AI Agents instead break my brain

    1 like

  • Kemal Ege Aktemur's profile picture

    @kemalegebuilds

    permissions drift and silent tool failures. agents rarely die in the prompt. they die when a tool returns almost-right data and nobody logged the handle.

  • Dr. Stephen Buscema's profile picture

    @DrBuscema

    for me its usually unclear ownership. agent does the draft fine then nobody owns the weird case so it all lands back on the founder

  • Jason Barrett's profile picture

    @JasonDigital

    Mine always break if I do no marketing.

  • Lea Thompson's profile picture

    @LeaT_Design

    breakage is the feature. my agents just don't talk to each other. like toddlers with nukes. shit.

  • Niraj Adhikari's profile picture

    @MINDFUEL_NIRAJ

    The quality of video is very good btw greg👀🫶🏻

  • Kanji Noguchi (野口寛士)'s profile picture

    @KanjiNoguchi

    For us it's almost never a 500 — it's a 200 with an empty array. We run a Japanese wholesale catalog that agents call directly, and queries land as short category nouns in whatever language the human typed; the agent passes that phrasing straight through instead of rewriting it for the index. When it misses, nothing errors, so the agent concludes the product doesn't exist and moves on. The other killer is latency shape: our product-detail reads sit around 20ms while search is two orders of magnitude slower, and search is the first call every agent makes. A slow front door in front of fast reads loses agents before they ever reach the useful part.

  • Alex D Kerner's profile picture

    @alexandrekerner

    @Abidebyabid Love the video, what did u use to make it?

  • Chinaza Joseph's profile picture

    @ChinazaJos10

    What breaks AI agents is relying on a single model without independent review. That’s why we use human-in-the-loop multi-agent orchestration: architecture, independent model review, and strict validation before execution.

  • Joulse's profile picture

    @joulsounet

    Long context plus a stale rule file is my usual failure mode

  • armorse's profile picture

    @MaR1236Cool

    What breaks agents for me: flaky tools + silent context rot. Curious what your top failure mode is when they look smart until they… don’t.

  • nomad.carpenter's profile picture

    @carpenter_17992

    Usually the tool loop—not the model: one bad retry policy can burn the budget while every individual call looks “successful.”

  • Relaystack's profile picture

    @Relaystack

    Bad instructions.

  • Nordin Sabli's profile picture

    @nordinbuilds

    Context collapse mid-task.