What is execution derailment in AI agents?
Execution derailment is the failure in which an LLM agent loop, as tool outputs pile up in its context, starts skipping steps, repeating tool calls, looping or inventing steps it never ran, and so fails to finish.
Google named the failure in its post on building ADK 2.0 (2026-07-01). In a typical agent loop every tool output is appended to the model's context. Over several iterations that adds prompt noise, and the agent becomes prone to loops, redundant tool calls and failing to finish. When the context gets crowded, the post says, the agent may skip steps or hallucinate execution paths.
The same pattern shows up after long pauses. Google's May 2026 guide to pause-and-resume agents observes that after long idle gaps a model often hallucinates intermediate steps that never happened, when it reconstructs progress from chat history.
Symptoms
Skipped steps
A required step, such as a verification or an approval check, is silently left out.
Repeated tool calls
The same lookup or action runs again because the earlier result is lost in the noise.
Loops
The agent cycles between the same few actions without progress.
Invented progress
The agent acts as if a step ran that never did, which is the most dangerous form when the next step has side effects.
Fixes
Move control flow into code
Google's answer in ADK 2.0 is a graph runtime: transitions are evaluated in code and only steps that need judgement use a model. In its refund example, tokens per run fell from 5,152 to 2,265 and latency from 7.2 to 5.7 seconds, figures Google labels illustrative, using mock APIs.
Keep state explicit
Track progress in a state machine or checklist stored outside the conversation, so the agent reads its current step instead of inferring it.
Shrink what each call sees
Pass each step only the data it needs, trim or summarise old tool results, and keep tool output terse.
Check before side effects
Verify that prerequisite steps actually completed before any irreversible action runs.
What it is confused with
Context rot
The underlying cause: recall degrades as context grows. Derailment is what that looks like in an agent's actions.
Hallucination
A broader term for invented content. Derailment is specific to the execution path of a multi-step loop.
Context anxiety
Anthropic's term for a model wrapping up early as it nears its context limit. That is stopping too soon, rather than wandering.
How Motiontic's engineers apply it
Motiontic's engineers look for derailment in traces, not just final answers: repeated tool calls, missing required steps and claims that do not match the log. Where a process has a known order, that order goes into code and the model handles only the steps that need judgement, which also limits what a prompt injection can redirect.
Execution derailment at a glance
| Symptom | Likely cause | Usual fix |
|---|---|---|
| Same tool called twice | Earlier result lost in a long context | Trim old results, keep tool output terse |
| Required step skipped | Control flow left to the model | Put fixed steps in code or a graph |
| Progress claimed that did not happen | State inferred from chat history | Store progress in explicit state |
| Agent cycles without finishing | No stop condition or budget | Iteration caps and done criteria |
Common questions
Why do AI agents skip steps?
As tool outputs accumulate, the context gets noisy and the model can lose track of which steps are done. Google calls this execution derailment and moves fixed steps into deterministic code.
Does a larger context window prevent execution derailment?
Not on its own. A larger window holds more noise as well as more signal. Explicit state, trimmed tool results and code-controlled steps address the cause.
Should every agent be turned into a graph to avoid derailment?
No. Google's own guidance keeps agents for ambiguous, subjective or unstructured work and uses workflows for predictable sequences. Many systems mix both.
Related
Context rot
Why agents get worse the longer they run, and what keeps context clean.
Agentic workflow
Model calls on a path set in code: predictable, cheaper and easier to test than a free-running agent.
Workflow vs Agent
A path fixed in code, compared with a model that chooses its own steps.
Agent graph
Nodes for model calls and code, edges for what runs next, and state that flows through.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.