An AI pipeline can report success while its agent repeatedly hits “File does not exist.” In a 2026 incident account, IT analyst Jula Markova traced at least 117 such errors to an ambiguous path convention in a brief-generation workflow. Retries often let the work finish, but they obscured the failures in run-level reports. The central lesson: normalize deterministic inputs in code and make the same path rule explicit to the agent.
What the 117 errors actually represent
Markova’s account, published on DEV Community on August 18, 2026, describes an audit of 1,811 session transcripts from May 15 through July 2. She found at least 117 instances of the same missing-file error. She calls that count a floor because transcript messages could be truncated, hiding additional instances.
As an Amazon Associate I earn from qualifying purchases.
That number is not an error rate. The denominator was transcripts, not path-resolution attempts, and the account does not establish how many attempts occurred. The figures are the author’s reported audit findings, not an independently verified measurement.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problems
File does not exist. Note: your current working directory is /Users/userxy/code/your-project.#1 Best Overall
The error line identifies a missing file and shows the current working directory, but neither by itself explains which path the agent tried or what base it used.
Why the same brief could lead to different paths
The workflow passed YAML briefs containing both absolute paths and repository-relative paths. The relative-path field had no declared base. As Markova explains, an agent could reasonably interpret it relative to its working directory, the repository root, or the directory containing the brief. Different interpretations could therefore point to different locations.
Markova’s shorthand for the mechanism is “Ambiguity in, probability out.” The input looked deterministic, but an unstated rule left a choice for the agent to infer. That is a contract-design problem, not evidence that every missing-file error has the same cause.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
How retries hid failures from run reports
According to Markova, a failed read was often retried: the agent tried another path and eventually found the file. The pipeline still delivered articles, and run reports recorded none of these failures, while the traces remained in session transcripts.
This creates two different views of reliability. A successful final outcome tells you that the task completed; it does not show whether the agent encountered errors, recovered through retries, or took a longer and less predictable route. Transcript-level review can reveal that hidden work, though it depends on what the transcripts retain.
What the audit found beyond path resolution
Markova also reported other classes of workflow problems in the commit accounting associated with the audit. These are her reported counts, not independent measurements:
| Issue reported | Count |
|---|---|
| Path-resolution failures | At least 117 |
| Writes to files never read first | 172 |
| Edits anchored to stale file content | About 20 |
| Sessions that dispatched an LLM fixer for a formatting error a script could fix deterministically | About 19 |
The post does not establish rates for these events or a shared cause for all of them. Together, they illustrate why auditing traces can uncover operational problems that a delivered result alone will not expose.
What changed in the reported fix
Markova says the team changed both sides of the path contract:
- Brief generator: It prefixed the worktree root so the generator could not emit a relative path.
- Agent contracts: The same convention was added to all five contracts: use absolute paths as provided and do not join them with the working directory.
The distinction matters. A prompt-only instruction asks the agent to interpret inputs consistently; producer-side normalization removes the ambiguity before the agent receives them. Applying the rule at both boundaries makes the expected behavior explicit and constrains what the producer can supply.
The rationale is broader than file paths: when a deterministic fact is already known by the producer, encode it in deterministic code rather than relying only on an agent to infer it. As Markova puts it, “Contracts for agents must be stricter than contracts for code.” That is her practitioner conclusion, not a formal standard.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What the verification does—and does not—show
Markova reports that a real generator verification produced 789 briefs with zero non-absolute paths, and that the TypeScript and Python test suites were green. The account does not provide the briefs or code change for independent inspection, so these remain reported verification results.
Free tools Windows power users keep installed
One-click scans. No signup required.
She also says transcripts were not mined again after the fix. The account therefore supports the narrower claim that the generator was changed so it could no longer produce a relative path under the stated convention; it does not establish that the error stopped appearing in later operation.
How to look for the same failure pattern
When an agent appears flaky, inspect the contract and the traces before assuming the model itself is the root cause. Useful questions include:
- Does the agent have to infer a path, identifier, ordering, format, or destination?
- Could a reasonable reader interpret an input in more than one way?
- Can retries turn an error into eventual success without surfacing the error in the run report?
- Are you reviewing only final outcomes, or also the retained session traces?
For any ambiguous input that can be resolved before the agent acts, define the convention and enforce it at the producer. Then state the matching rule in the consumer contract. Monitor outcomes and inspect traces as complementary signals: successful completion does not erase the operational significance of hidden retries.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
