Topic starter
20/08/2026 6:10 pm
TL;DR: Production failures cluster around retrieval noise, hallucinated tool calls, recursive loops, guardrail gaps, and schema drift, with traceability and trajectory-level evaluation emerging as the core operational fix, according to Arize field analysis of millions of agent decision paths. The central issue is not that agents are merely inaccurate, but that deterministic enterprise workflows now absorb probabilistic behaviour that can silently fail, misfire, or mask errors.
Editorial analysis by NHI Mgmt Group, based on content published by Arize: “Why AI Agents Break: A Field Analysis of Production Failures”.
Key questions
Q: What breaks when AI agents are monitored like ordinary automated jobs?
A: What breaks is the assumption that fixed schedules and static logs are enough.
Q: Why do AI agents create more risk once they move from retrieval to tool calling?
A: Retrieval only returns information, but tool calling lets an agent interact with live business systems and change records, approve actions, or trigger workflows.
Q: How do security teams tell whether an agent is actually using the right context?
A: They should check the specific spans or chunks that influenced the final answer, not just whether the right document was retrieved.
Practitioner guidance
- Implement trajectory-level monitoring Map agent decisions from prompt to tool call to backend effect so you can see loops, guesses, and silent failures instead of only final status codes.
- Enforce deterministic runtime guardrails Block destructive or policy-violating outputs at the tool boundary, where the agent is about to act, rather than relying on prompt wording alone.
- Track span-level evidence usage Measure which text chunks were actually used in the final logic, especially for retrieval-heavy workflows, because document presence does not prove decision quality.
Bottom line: AI agents in production fail in repeatable ways, especially when retrieval, tool calls, and long contexts interact with rigid business workflows.
Explore further
View Full Forum → | NHI Foundation Course → | Our Services → | Read the full analysis →
20/08/2026 6:46 pm
Trajectory-level observability is now the minimum viable control for agentic AI. Linear logs answer the wrong question because they show status, not decision shape. When an agent succeeds inefficiently, guesses through schema drift, or loops through the same tool call, the failure only becomes visible in the path. Practitioners should treat trajectory analysis as a control layer, not a reporting feature.
A few things that frame the scale:
- 70% of organisations grant AI systems more access than they would give a human employee performing the exact same job, according to the 2026 Infrastructure Identity Survey.
- 19% of organisations give AI systems dramatically more access than human employees, nearly one in five granting unrestricted privilege, according to the 2026 Infrastructure Identity Survey.
A question worth separating out:
Q: What should organisations do when an agent keeps behaving correctly but inefficiently?
A: They should treat repeated calls, tight loops, and long trajectories as a performance and control issue, not just an optimisation problem. An agent that eventually succeeds can still become commercially or operationally unsafe if it burns tokens, repeats work, or amplifies backend load. The right fix is path analysis plus runtime limits.
👉 Read our full editorial: Why AI agents break in production and what that means