Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
When an AI agent appears stuck, do not start another blind retry. Freeze the run, preserve its trace, and check four signals together: whether its state is changing, whether the same tool call and arguments recur, whether token use is climbing, and whether elapsed time is abnormal. No single signal proves a loop, but their combination is a strong diagnosis. The fix is usually a bounded execution policy plus a correction to the tool, state update, verifier, or prompt that prevents progress. This guide shows how to identify the loop, capture evidence, stop it safely, repair the cause, and prove the repair works.
What does a stuck or looping AI agent look like?
A stuck agent is not merely slow. It is an execution whose actions no longer move the task toward a verifiable goal. Camunda cautions that no single health indicator proves an agent is stuck; confidence comes from several signals occurring at once.
As an Amazon Associate I earn from qualifying purchases.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errors| Signal | What to inspect | Why it matters |
|---|---|---|
| Unchanged state | Files, database rows, workflow status, memory, or task variables remain identical between iterations. | An action that produces no new state gives the next decision no new information. |
| Repeated action | The same tool and nearly identical normalized arguments recur, or two calls alternate in a cycle. | Repetition reveals that the policy is selecting an action it has already tried without benefit. |
| Growing usage | Tokens, model calls, retries, or tool invocations increase without a corresponding state transition. | Cost and context grow even though the task is not advancing. |
| Disproportionate duration | Wall-clock time or latency is far beyond the normal range for this task. | A long run can be legitimate, but abnormal duration combined with the other signals indicates a stall. |
| No completion signal | The verifier never returns success, or the agent cannot express a bounded failure. | Without a reachable stop condition, the model has no reason to end. |
A run can loop with a zero error rate. Amazon Web Services showed a 2026 CloudWatch example with 266.9K total tokens and a 0% error rate, illustrating that successful API responses can still form an unproductive cycle. Another 2026 AWS example contained 177 spans and approximately 85,590 milliseconds of average latency as evidence of a loop. Treat telemetry as behavioral evidence, not just an error counter.
What should an agent trace contain?
Start with a complete execution record rather than a new retry. Databricks describes a trace as a structured record of the complete execution, including intermediate language-model calls, retrievals, tool invocations, latency, and errors. For debugging, add the state and decision context that explain why each action was chosen.
#1 Best Overall
- [Ultra-Lightweight Ear Buds Designed for Small Ears] Each earbud weighs only 3.7g and features a compact, ergonomic in-ear design made especially for small ears. Secure, low-profile, and comfortable for workouts, all-day wear or overnight sleeping.The ultra-slim profile sits flush in your ear with no hard edges poking or rubbing, so you can even sleep on your side without discomfort.
- [Immersive Stereo Sound with TOZO OrigX Technology] TOZO OrigX tuning delivers clear vocals, balanced mids, and natural stereo sound for music, podcasts, and videos.
- [Long Battery Life for Daily Use] Get up to 7 hours of playtime on a single charge, with up to 32 hours total using the charging case—ideal for workdays, commuting, and extended listening sessions.
- [Bluetooth 5.3 & Stable Connection] Bluetooth 5.3 provides fast pairing, stable wireless performance, and reduced dropouts as you move around home or office.
- [Deep Bass with Clear Vocals] High-performance drivers produce punchy bass while keeping vocals clean and detailed for everyday listening.
Capture the request and configuration
- The user goal and any constraints or authorization context.
- Agent, model, system prompt, tool descriptions, prompt version, and configuration values.
- Conversation or task identifiers, tenant or environment, and a correlation ID that follows every downstream call.
Capture every decision and effect
- Each model request and response, including token usage, finish reason, and timestamp.
- The selected tool, exact arguments, normalized arguments for comparison, result or error, HTTP status, and latency.
- Retrieval queries and returned document identifiers when retrieval is involved.
- Memory reads and writes, state before and after the action, and a compact diff of changed fields or files.
- Verifier input and output, retry count, iteration number, escalation reason, and final stop reason.
JetBrains specifically recommends logging iteration count, selected action, exact tool inputs and outputs, state changes, retry count, stop reason, and escalation reason. Google Cloud lists prompts, responses, token usage, latency, error rates, and tool or API exchanges as core agent-observability signals. Redact secrets and personal data before storing traces, but do not remove the identifiers needed to correlate an action with its result.
Use a normalized action signature
Raw JSON can differ in harmless ways, such as key order or a timestamp. Create a stable signature from the tool name and arguments after removing volatile fields and sorting keys. Keep both the original arguments for forensics and the normalized signature for loop detection. Record a hash only as an index; retain the readable values subject to your security policy.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
signature = hash(tool_name + canonicalize(args, remove=['request_id', 'timestamp']))
if signature == previous_signature and state_hash == previous_state_hash:
repeated_attempts += 1Also detect short cycles. A sequence such as read_status, write_update, read_status, write_update may be a loop even when no single call appears three times consecutively.
How do you triage a running agent safely?
Use this order so you preserve evidence before changing the conditions that produced the failure.
- Freeze the run. Cancel the model stream or worker, disable further side-effecting tools, and preserve the current external state. If cancellation is unsafe, revoke only the permission that can cause additional damage and let a read-only drain complete.
- Check for a pending human step. VS Code advises looking for an unanswered question, approval request, or command that is still running. A waiting approval can look like a model loop when the agent has not actually received the next observation.
- Locate the first divergence. Review the last known-good state, recent tool calls, terminal output, changed files, and the earliest point where the execution differed from an expected trace. The first wrong action is usually more useful than the final repeated call.
- Group actions by tool and normalized arguments. Count identical calls, alternating cycles, retries, and calls whose outputs are byte-for-byte or semantically unchanged.
- Compare observations with state. If a tool returns the same result while the task state remains unchanged, the retry is not productive. If a tool reports success but the expected state did not change, investigate stale reads, denied writes, transaction rollback, caching, or idempotency behavior.
- Check routing and tool contracts. Determine whether the selected tool can perform the requested operation. AWS identifies incorrect tool selection as a common cause of loops; a model may repeatedly call a search tool when it needs a write tool, or invoke a write tool without the required precondition.
- Test the completion predicate. Run the verifier against a known-complete state and a known-incomplete state. If it returns the same result for both, or can never return success because a field is never populated, the agent has no reachable stop signal.
Do not delete the trace after restarting. Preserve the stopped run, the external state snapshot, and the exact intervention so you can distinguish a real fix from a lucky completion.
Why do agents repeat actions, and how do you fix each cause?
No termination condition or an unreachable verifier
An agent needs an explicit completion predicate and a bounded failure path. Microsoft’s Agent Framework guidance is direct: “Always bound autonomous loops.” Define what success means in observable state, then define what happens when success cannot be achieved within the allowed budget.
Recommended Free Tools
Rank #2
- 【IP68 Waterproof Multifunctinal Smart Watch】Torjalph fitness activity tracker smart watch is IP68 waterproof technology ,you can wear it when swimming,handwashing,it's rain-proof which means that you perform multiple exercise modes outside without worrying about sweat or rain corroding your fitness smart watch,and this watch equipped with a 230 mAh large capacity battery magnetic charging ,and this sports activity tracker watch can be used 5-7 days If fully charged.
- 【Larger 1.83'' Full Screen Watch & Multi-dial Selection】1.83 inch full HD color touch screen smart watch for men / women provides excellent 240*284 resolution image quality ,This Bluetooth smart watch offers a better touch experience and remarkable immersion and ,more than 200 watch faces for you to choose from App"Da Fit ", you can change the dial style according to your mood to wear keeping the world on your wrist
- 【Accurate Activity Tracker Smart Watch 】sports watch activity tracker uses sweat and dust resistant breathable silicone strap, our Fitness Tracker Smart Watch for Android iOS Phones supports 12 sports modes,walking,running,bicycle,swimming,badminton,basketball,football,rope skipping. This bluetooth calls smart watch for men women is lightweight and stylish, bringing more convenient and relaxing sports experience while also providing more protection.
- 【Your 24 Hour Personal Daily Health Monitor】Fitness watch for women men full touch screen IP67 waterproof bluetooth smartwatch for Android iOS Phones with heart rate blood pressure monitor multifunction Andriod smart watch with steps counter sleep monitor sports fitness watch, monitoring blood oxygen saturation allows you to understand your physical condition in real time to track fitness progress ,to know your health status in time.
- 【More Functions Bluetooth Call Smart Watch】This fitness activity smart watch full screen for iPhone and Android phones includes a bunch of new features including:The bluetooth smart watch can hang up the phone, receive and display WhatsAPP /Text /Facebook /Twitter /Messenger reminder.Support sedentary reminder, alarm clock, stopwatch, timer,weather information,music control, remote camera control compatible with Android 5.0 and above or iOS 9.0 and above smartphones.
For example, a deployment task might require status == 'deployed' and a matching version identifier. A failed path should return blocked with the last successful state and the missing precondition, rather than asking the model to continue indefinitely. Test the predicate independently before putting it in the loop.
Repeated-action blindness
Keep a per-run history of action signatures and state hashes. After a small number of unsuccessful identical actions, stop or force a replan. AWS recommends forcing termination after three identical repeated actions. That is a practical default for a first guardrail, not a universal number: a polling operation may legitimately repeat when its state is expected to change, so pair the count with a maximum polling interval and evidence that the observed value is changing.
Wrong tool or ambiguous tool description
Improve each tool contract with its intended use, exclusions, required arguments, side effects, failure modes, and a short example. State when the tool must not be used. Add schemas that reject impossible combinations before execution, and include structured error messages that tell the agent whether to correct arguments, choose another tool, or escalate. Then verify the selected tool in traces; a better description is not a fix if routing still chooses the wrong capability.
State is not updated or observations are misleading
Return structured observations that clearly identify what changed. A write tool should report the resulting version, resource identifier, and durable status, not only “request accepted.” A read tool should include freshness or revision information when stale data is possible. Make the next-step policy depend on those fields, and fail closed when the expected state transition is absent.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Retries add no information
Separate transient failures from deterministic failures. Use bounded retries with exponential backoff for timeouts, rate limits, and temporary network errors. For a validation error, permission denial, schema violation, or reproducible business-rule failure, change the strategy or escalate instead of repeating the same request. Every retry should have a reason code and should record whether it produced new information.
The prompt says never to give up
Replace open-ended instructions with explicit stop rules. A useful policy is: after three unsuccessful identical actions, explain the blocker, preserve the last successful state, and ask for human help or select a documented fallback. Phrase persistence as trying a bounded set of alternatives, not as continuing forever.
Rank #3
- [Low Noise High Fidelity] New bee guitar cord with 22AWG oxygen-free-copper center conductors provides enhanced signal transfer, the PVC insulation material and the high-density stranded copper double shield provide noise rejection for quiet operation
- [Ideal for Choice] New bee 1/4 straight and right angle metal connectors, ensures a secure connection in any instrument, pedal, or amp jack. L shape great fro on stage guitar stands
- [Wide Compatible] Suitable for mixer, power amplifier, microphone, effector, decoder, equalizer, electoronic organ, electric guitar, electric bass and other electronic equipments. Widely used in the installation of smart home, karaoke speaker, professional stage audio, etc
- [Easy Disassemble] Making incidental repairs quick and easy! Very nice, flexible cable
- [Dont Hesite to Order] 24 hours customer services and professional technology team are standing by
External effects are hidden, delayed, or non-idempotent
An agent may repeat a payment, ticket update, or file operation because the first request succeeded but its confirmation was lost. Use idempotency keys for side-effecting operations, expose the server’s operation status, and make a follow-up read safe. If the effect cannot be made idempotent, require approval before a repeat and route uncertainty to a human.
Which runtime guardrails should every autonomous loop use?
Use several independent limits. A single token limit will not stop a fast loop with cheap calls, and a time limit will not protect a slow tool that consumes little context.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →| Guardrail | What it limits | Design notes |
|---|---|---|
| Maximum iterations | Number of model-decision cycles. | Count completed and failed attempts; include a reason when the cap trips. |
| Wall-clock deadline | Total elapsed execution time. | Set a deadline appropriate to the workflow and reserve time for cleanup or escalation. |
| Token budget | Prompt and completion consumption. | Track per run and per phase; a token cap is a cost guard, not proof of progress. |
| Per-tool retry limit | Attempts against one tool or endpoint. | Use separate limits for transient retries and identical-action repeats. |
| Stale-state limit | Consecutive iterations without a meaningful state diff. | Compare semantic state, not timestamps or request IDs that change every call. |
| Human-escalation condition | Risky or ambiguous situations. | Escalate on permission uncertainty, irreversible effects, conflicting observations, or exhausted alternatives. |
AWS gives example conversational-agent budgets of 5,000–10,000 tokens per session and 10–15 reasoning steps. Those are documented examples to tune against your workload, not universal defaults. Microsoft’s LoopAgent example uses a completion evaluator together with configurable MaxIterations; the important pattern is a reachable evaluator plus a hard ceiling.
Return a structured guard-trip result
When a bound trips, return enough information for an operator or supervisor agent to act without rereading an entire log.
{
'status': 'blocked',
'stop_reason': 'identical_action_limit',
'last_successful_state': {'job': 'build-42', 'status': 'queued'},
'repeated_action': {'tool': 'check_job', 'signature': 'a91f...'},
'attempts': 3,
'elapsed_ms': 18400,
'suggested_intervention': 'verify worker availability or choose another queue',
'trace_id': 'trace-7f2c'
}Persist the trace and the structured result together. A supervisor can then replan with a different tool, request approval, or end the task without guessing what happened.
How can you tell whether a long-running agent is making progress?
Measure progress as state transitions, not activity. A legitimate long run may issue many calls while waiting for an external job, but each polling cycle should update a version, status, timestamp with a documented freshness guarantee, or other meaningful observation.
Outdated 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 matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Rank #4
- Digital Family Calendar: BSIMB digital calendar wall touch screen puts an end to your chaotic life. Designed for busy families, the digital calendar keeps you organised and in step with your family. Assign different colours and tasks to each family member to make family scheduling easier and more intuitive.
- Easy Setup and Auto-Sync: Plug in, connect to WLAN and auto-syn your electric calendar in minutes. Digital wall calendar compatible with Google calendar, iCloud calendar, Outlook calendar, Cozi calendar and Yahoo calendar. Add events directly on the device or through the ecalendar app. Electronic calendar stays current whether you're at work or traveling—no manual updates needed.
- All-In-One Home Assistant: Smart calendar builds habits and encourage independence with personalised chores. Display daily meal plan to reduce the daily ‘What's for dinner?’ stress. To-do lists keep track of everything you want. Turning chores into valuable achievements with stars and rewards, making life easier and more enjoyable.
- Manage Anytime Anywhere: BSIMB electronic calendar for family stay connected with the powerful ecalendar app. Access and manage your digital wall calendar anytime, anywhere with free features like stars & rewards, photo screensaver, chore charts, meals plans, to-do lists etc. Electronic planner paid features like Magic Import tailor advanced tools for users with special needs.
- Stay in Touch with Loved Ones: Share photos or videos from your phone to electronic wall calendar with app, instantly update your life moment, and keep in touch with your family, especially those who are far away; you can also surprise your smart calendar digital wall by sending him a wish card; it's also great for keeping tech-illiterate elders in touch with the youngsters.
Define milestones before execution
- List the expected states in order, such as
queued→running→validated→complete. - Specify which tool is allowed to cause each transition and what evidence confirms it.
- Set a maximum dwell time for each state and a fallback when that time expires.
Use a progress ledger
Record the current milestone, last meaningful state change, next permitted action, and deadline. If the agent calls tools but the ledger has not changed by the stale-state threshold, classify the run as stalled even if every call returned HTTP 200. This also distinguishes waiting for an external system from a reasoning loop: waiting has a changing deadline and a documented poll condition; a loop repeats without either.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How should you verify that a loop fix really works?
Do not declare success because one replay completed. Databricks describes a repeatable quality cycle: trace executions, find an issue, curate representative failures into an evaluation or golden dataset, write a scorer, fix the agent, rerun evaluation, and monitor production.
Build a regression case from the failure
- Save the original goal, prompts and tool versions, exact looping sequence, external-state snapshot, and stop reason.
- Describe the expected state transitions and the maximum permitted repeat count.
- Define the desired escalation message, including the blocker and last successful state.
- Replay the case with deterministic tool fixtures where possible, then run it against a safe staging service.
Compare before and after
Score completion rate, repeated-call count, total steps, latency, token use, guard trips, escalation quality, and unintended side effects. A lower step count is not automatically better if it hides a failed write; require the expected final state and an auditable stop reason. Include near-miss cases: transient timeout, stale read, permission denial, malformed arguments, and a tool that legitimately needs several polls.
Monitor production for recurrence
Alert on clusters of identical signatures, rising stale-state time, unusual token-to-state-change ratios, and guard-trip rates by agent version and tool. Sample traces for human review, annotate confirmed loops, and feed those examples back into the evaluation set. Roll back a prompt, tool schema, or model change when loop indicators rise after deployment.
How do observability approaches differ?
Choose an approach by the evidence and controls it provides, not by the volume of dashboards. The same comparison applies whether you use a vendor platform, cloud-native telemetry, or an internal event store.
| Decision axis | Questions to ask | Minimum acceptable capability |
|---|---|---|
| Trace completeness | Can you see model, retrieval, tool, memory, verifier, and state-change spans in one execution? | Every action has exact inputs, outputs, latency, and a correlation ID. |
| Detection controls | Can the system hash repeated calls, detect cycles, enforce step and token limits, and alert on stale state? | Automated thresholds plus a way to inspect the triggering evidence. |
| Debug workflow | Can engineers query, replay, annotate, compare versions, and turn failures into evaluation cases? | Searchable traces and exportable fixtures with redaction controls. |
| Operational fit | Does it meet cloud integration, data-residency, access-control, cost, and latency requirements? | Documented retention, least-privilege access, and measured telemetry overhead. |
Databricks positions tracing as the foundation for agent observability and quality. AWS documents tracing and loop evidence through Bedrock AgentCore and CloudWatch examples, while Google Cloud documents agent-observability signals such as prompts, responses, tokens, latency, errors, and tool exchanges. These products differ in deployment and governance; the diagnostic fields above are the portable requirement.
What should an on-call runbook contain?
Keep the response short enough to use under pressure, but require evidence at each decision.
- Stop: identify the run, freeze side effects, and preserve the trace and external-state snapshot.
- Classify: mark pending approval, running command, external wait, deterministic failure, or suspected loop.
- Measure: inspect state diffs, normalized action counts, token growth, latency, and verifier output.
- Contain: apply the iteration, time, token, retry, and escalation bounds; revoke risky permissions if needed.
- Correct: fix the first divergence—tool routing, contract, state update, idempotency, verifier, or stop rule.
- Replay: run the saved regression case and confirm the expected final state and escalation behavior.
- Learn: add the case to evaluation, annotate the trace, and monitor the deployed version for recurrence.
The objective is not merely to make one run finish. It is to make future failure observable, bounded, and recoverable.
Free tools Windows power users keep installed
One-click scans. No signup required.
Frequently Asked Questions
Can an AI agent be stuck even when every tool call succeeds?
Yes. A tool can return a valid response that leaves the task state unchanged, so the agent keeps receiving successful but useless observations. Check state transitions and verifier outcomes alongside HTTP or API error rates.
Should I raise the token or iteration limit when an agent loops?
Usually not as a first response. A larger budget can increase cost and side effects while hiding the missing stop condition; first determine whether the next action can produce new information, then tune limits from measured successful runs.
How do I distinguish polling from a reasoning loop?
Polling has an explicit condition, a bounded deadline, and evidence that the observed revision or status can change. A reasoning loop repeats the same action without a meaningful state diff or a new permitted transition.
What is the safest way to replay a looping agent?
Replay against recorded or simulated tool responses first, with side-effecting tools disabled. When a live test is necessary, use a staging environment, fresh idempotency keys, least-privilege credentials, and the same guardrails as production.
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.
