AI agents fail for more than model errors
Principal Cloud Solutions Architect
John Savill examines failure points around AI agents, including context, tools, state and security, giving teams more than model quality to investigate.
Key insights
- The video asks whether AI agent failures are really the model's fault.
- Its chapter list covers fuzzy goals, bad context, tool misfires and bad reasoning.
- It also names lost state, excessive privileges and limited visibility as failure topics.
- The video includes an improvement loop and two caveats, but the supplied description does not explain them.
John Savill's [MVP] published a YouTube video titled “Why do AI Agents Fail,” organized around the question of whether failures are all the model's fault. The supplied description identifies topics including goals, context, tools, reasoning, state, permissions and visibility. The central takeaway is that investigating an agent means looking beyond its model, while the available description does not provide enough detail to attribute particular failures or prescribe fixes.
What's new
The video frames agent failure as a set of possible problems across the system surrounding the model. Its chapter list moves from “the model and what surrounds it” to specific topics: fuzzy goals, bad context, tool misfires, bad reasoning and lost state. It also names “hijack and over privileged” and “flying blind,” followed by an improvement loop, two caveats and a summary.
That structure is useful as a diagnostic checklist, not as a claim that every agent will encounter each issue. The description does not define the terms, show a particular agent configuration or explain how Savill evaluates the causes. Savill also provides a whiteboard link alongside the video, which is listed as a source below.
Why it matters
For teams building or operating agents, the question “Is it the model?” can be too narrow. The chapter topics suggest checking whether the task is clearly specified, whether the agent has appropriate context, whether its tools behave as expected and whether it retains state. In practical terms, that means treating the agent and its surrounding components as things to examine, rather than assuming a model change is the answer.
Admins and security teams may also find the listed topics “hijack and over privileged” relevant. They point to concerns around hostile influence and excessive access, although the supplied material does not say what safeguards Savill recommends or how he demonstrates these concerns. Makers and developers should read the video as a prompt to assess the design and operation of their own agents, not as product-specific guidance.
The material does not identify a Microsoft product, licensing requirement, rollout stage or feature announcement. It therefore does not establish a change to Copilot Studio or another agent platform. Its value for Microsoft 365 and Power Platform professionals is the broader systems perspective, not instructions for a named product.
How it works
The description presents a sequence of topics rather than a step-by-step procedure. Based on that sequence, a team reviewing an agent can use these areas as questions to investigate; this is a way to apply the chapter headings, not a procedure quoted from the video.
- Check the goal: Decide whether the request is specific enough to guide the agent. The chapter calls this a “fuzzy goal,” but the description gives no example.
- Review the inputs: Examine whether the agent has the context needed for its task. The chapter label is “bad context”; no particular context source or remedy is specified.
- Inspect actions and reasoning: Consider whether tools misfire or the agent reasons poorly. The description identifies both as topics but does not give a tool, transcript or troubleshooting method.
- Check continuity and controls: Look for lost state, possible hijacking, excessive privileges and limited visibility. The chapter titles flag these as areas of concern without defining their mechanics.
- Consider improvement and caveats: The video includes an improvement loop and two caveats, but their content is not available in the supplied description.
This framing can help teams avoid collapsing every failure into “the model got it wrong.” It does not, by itself, identify a root cause: the chapter list supplies categories, not evidence about a particular incident.
Limitations and open questions
The source material available here is a description and chapter list, not a full transcript. It does not tell us what examples Savill uses, what the two caveats are, how the improvement loop works or what concrete mitigations he recommends. We therefore cannot verify any detailed explanation of tool behavior, permissions, state handling or agent security from this material.
The description also does not name an agent platform or discuss licensing, availability or deployment requirements. Readers looking for a product setup guide, a security control checklist or measured comparisons between models will need information beyond the supplied description. The linked whiteboard may provide additional context, but its details are not reproduced here.
Sources
Frequently asked questions
Does the video say AI agent failures are always the model's fault?
No. Its description asks whether they are all the model's fault and lists issues involving goals, context, tools, reasoning, state and security.
Does the supplied material explain how to fix these failures?
Not in detail. The chapter list mentions an improvement loop, but the provided description does not explain its steps or specific fixes.
Is this a guide to a particular Microsoft agent product?
The supplied description does not name a Microsoft product, so it does not establish product-specific setup, licensing or availability guidance.