Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOutdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchReliable AI agents need more than better prompts. In an account published on August 29, 2026, software engineer Tamiz Uddin describes 96 refusals or failed iterations before what he calls a successful 97th attempt. His central lesson: reliability comes from engineering the system around a stochastic model—handling uncertainty, tool failures, evaluation, tracing, and escalation—not from prompt changes alone.
What the 96-refusal story does—and does not—show
Uddin presents the figure as a personal engineering account, not as a controlled study. The article shifts among “refusals,” “iterations,” “failed deployments,” and “attempts,” without defining them as equivalent units. Treat the number as his narrative of repeated failures before success, not as a validated incident count or an industry benchmark.
As an Amazon Associate I earn from qualifying purchases.
The article also gives outcome figures without an underlying dataset or a detailed evaluation method. They may illustrate how Uddin says his system changed; they do not establish that the same changes will produce the same results elsewhere.
Recommended Free Tools
Why refusals deserve diagnosis, not just another prompt tweak
A refusal can be the right outcome, an overbroad policy response, a sign that the agent lacks needed context, or a reaction to an ambiguous request. Uddin’s taxonomy separates those cases so teams can choose a targeted fix rather than treating every refusal as the same defect.
| Category | What it suggests | Possible response |
|---|---|---|
| Legitimate refusal | The request should not be fulfilled under the applicable policy. | Keep the refusal path and verify that it is clear and appropriate. |
| Overrefusal | The agent declined a request it could have handled. | Review policy wording and add relevant domain context. |
| Context gap | Necessary information was missing from the request or available context. | Ask a focused clarification question or retrieve the missing information. |
| Ambiguity | The request could reasonably mean more than one thing. | Clarify intent before taking an consequential action or giving a definitive response. |
According to Uddin’s account, the refusal categories were 31% legitimate refusals, 47% overrefusals, 14% context gaps, and 8% ambiguity. He also reports that adding domain-specific context reduced overrefusals by 62%. The article does not provide the sample size or classification method behind these figures, so they should be read as his reported experience rather than general rates.
Put an uncertainty gate before a confident answer
Uddin recommends giving the agent a way to signal that it lacks enough confidence or context, then request missing information rather than force an answer. This is a system behavior, not simply a more cautious sentence in a prompt: define what uncertainty means for the task, what information can resolve it, and what the agent should do if it cannot resolve it.
Uddin reports that adding an uncertainty gate eliminated 60% of production incidents. No independent incident definition or measurement method is supplied, so the figure is specific to his account. The practical lesson is to measure whether the gate catches known failure cases and whether it creates unnecessary interruptions, rather than assuming that a confidence signal is automatically reliable.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Make tool failures bounded and recoverable
External tools introduce failure modes a model prompt cannot fix: a service may be slow, unavailable, or return unusable data. Uddin’s approach is to define the recovery behavior explicitly instead of allowing an agent to retry indefinitely or conceal a failure.
- Set a timeout for each tool call so a stalled dependency cannot hold a task open without limit.
- Use a bounded retry count, with an appropriate delay where retries are useful.
- Specify a fallback step for a failed call, such as using an alternate source or asking a person for the missing input.
- Allow graceful degradation when a nonessential capability is unavailable, while making clear what the agent could not do.
- Record whether a retry or fallback was used and what result it produced.
The exact limits depend on the task and dependency; Uddin’s article does not prescribe universal timeout or retry values. A fallback should not silently turn missing evidence into a confident answer.
Evaluate the whole system under production-like conditions
A model can perform well in an isolated test and still fail when the surrounding system is busy or a dependency behaves differently. Uddin recommends testing the full agent workflow under conditions that resemble production, including:
- Latency variation, rather than only fast, predictable responses.
- Concurrent requests, to expose contention and coordination problems.
- Cold caches, where previously stored results are unavailable.
- Dependency failures, such as a tool timing out or returning an error.
These tests help reveal whether timeouts, retries, fallbacks, and user-facing error paths work together. They do not by themselves prove safety or reliability; teams still need task-specific acceptance criteria and review of the resulting behavior.
Trace decisions, not only final outputs
An output-only log can show what the agent said without explaining how it arrived there. Uddin argues for tracing decision points so engineers can investigate why an outcome occurred. For an agent workflow, useful records may include the decision stage, confidence or uncertainty signal, tool calls and results, retry history, selected fallback, and whether a person was asked to intervene.
Uddin describes this as a way to debug causes rather than just inspect responses. In practice, traces should be designed around the questions a team needs to answer, while limiting sensitive data and access according to the system’s privacy and security requirements.
Escalate based on uncertainty and impact
Sending every ambiguous case to a person can make human review a bottleneck; letting every uncertain case proceed can expose users or the business to avoidable harm. Uddin proposes considering both uncertainty and business impact: higher uncertainty combined with higher impact is a stronger reason to escalate than uncertainty alone.
According to Uddin, risk-based escalation reduced human intervention by 85%. The account does not define the intervention measure or establish that the approach is appropriate for high-stakes decisions. Escalation criteria should reflect the consequences of error in the particular application, and the article’s figure should not be treated as safety evidence.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Use effort progressively instead of paying maximum cost on every task
Uddin’s reported metrics suggest a system that can succeed at different effort levels, with more work available for cases that need it. That design idea is distinct from simply repeating the same call: a later stage might request clarification, use another tool, apply a more thorough verification step, or route the case for human review. Each additional stage needs a stopping condition so cost and latency do not grow without bound.
Uddin reports the following before-and-after figures for his system. The article does not supply the measurement definitions or test conditions, and the values are not independently verified.
| Measure | Earlier figure | Later figure |
|---|---|---|
| Success rate | 68% | 96%+ |
| Average cost per success | $0.31 | $0.047 |
| Human escalation | 34% | 2.1% |
| P99 latency | 4.2 seconds | 6.8 seconds |
He separately reports 87.3% first-attempt success, resolution within five attempts for 96.1% of cases, and resolution within 100 attempts for 99.2%, alongside a mean cost per resolution of $0.047 and mean latency of 6.8 seconds. These are figures from Uddin’s account, not a reproducible benchmark; the article does not establish how “attempt,” “resolution,” or the latency population were defined. The longer P99 latency in the before-and-after table also illustrates a trade-off: a system that spends more effort recovering may improve its reported success rate while taking longer at the tail.
A practical review checklist
- Classify refusals as legitimate, overrefusal, context gap, or ambiguity before changing prompts or policy.
- Define what the agent should do when it is uncertain, including when to ask for clarification.
- Set timeouts and bounded retries for each external dependency, with explicit fallback behavior.
- Test concurrency, slow responses, cold caches, and dependency failures across the full workflow.
- Log decision traces that explain tool use, uncertainty, retries, fallback, and escalation.
- Set human escalation criteria using both uncertainty and the consequences of a wrong result.
- Track success, cost, latency, refusals, and intervention together so improvements in one measure do not hide regressions in another.
Uddin says the process took roughly 14 weeks from first deployment to production stability. That timeline belongs to his project; it is not a general estimate for other teams or a guarantee that this checklist establishes production readiness.
Source and scope
This article summarizes Tamiz Uddin’s account, republished on DEV Community on August 29, 2026. The republication links to his original article at tamiz.pro. The reported outcomes above are attributed to Uddin because the article supplies no independent study or detailed evaluation methodology.
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.
