agent failed to resume after 429 backoff, scan abandoned mid-tenant

Not specified by the original author.

TL;DR

Persist the pagination cursor before backing off, and resume from it after the wait. The scan is abandoned because the agent loses its place during the backoff.

Error

(scan stopped at 40 percent after 429 responses; no resume attempted; no error text)

Steps

1. Confirm the abandonment: the agent log shows 429s followed by exit, with no resume. Expected: the failure mode confirmed.
2. Add cursor persistence: write the current page token or offset to durable state before sleeping for backoff. Expected: the place saved.
3. On wake, read the cursor and continue the scan from there instead of restarting. Expected: resume logic in place.
4. Re-run a throttled scan. Expected: it pauses on 429s and completes.

When to use

• Scans stop partway after 429 responses
• The agent has backoff but no resume
• Large tenants always trigger throttling

When not to use

• The scan fails without any 429 (a different error)
• The API has no pagination cursor (page by offset or filter instead)
• Throttling never happens (the resume logic is still good hygiene)

Tool compatibility

Any paginated license API (Okta, Graph, Salesforce, Slack).

Variant phrasings

agent restarts from page one after backoff

That re-fetches and re-throttles. Persist the cursor to avoid it.

cursor lost when the agent process restarts

Write the cursor to disk or a state store, not just memory.

Why it happens

Backoff without state is just a pause before giving up. The scan's position lives in memory, so anything that interrupts the run (exit, restart, timeout) loses it.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_MUhH90JKJK4A9Q6KmHPvIw

Resolved from the public thread: https://vectle.com/posts/pst_MUhH90JKJK4A9Q6KmHPvIw

Thread query: agent failed to resume after 429 backoff, scan abandoned mid-tenant

TL;DR: Persist the pagination cursor before backing off, and resume from it after the wait. The scan is abandoned because the agent loses its place during the backoff.

Not specified by the original author.

TL;DR

Back off properly instead of retrying into the limit, because every fast retry extends the penalty and risks an abuse block. Read the retry guidance in the Microsoft Entra ID Graph response, slow the agent to a paced rate with jitter, reduce page sizes and parallel calls for this step, and checkpoint after every batch so the run can pause until the window resets and then finish. A rate limited review that checkpoints loses minutes; one that restarts loses the whole run and hits the limit again at the same place.

Entra ID Graph API rate limit hit mid-campaign-status-poll: agent failed with error 429 too many requests

Use this when

• The run fails with a 429 response from Microsoft Entra ID Graph partway through the work to check campaign completion.
• Bursts of parallel agent calls succeed briefly and then all start failing together.
• The same run succeeds overnight or at lower concurrency.

Not for this skill when

• Calls fail with authorization errors at any rate. That is permissions, not pacing.
• A single call is rejected as malformed. Fix the request shape instead of the rate.

Steps

1. Capture a failing Microsoft Entra ID Graph response in full, including its status, any retry after guidance and remaining quota headers, and record how far the run had progressed. Verify: the limit details and the progress point are recorded from a real response.
2. Stop the run cleanly and save a checkpoint of completed batches so the pause does not discard finished work. Verify: the checkpoint lists completed batches with their scopes.
3. Reduce the step concurrency to a single paced worker for this API and set delays between calls with random jitter, sized from the quota the platform documents for this endpoint family. Verify: the pacing configuration is set and differs from the failing configuration.
4. Shrink page or batch sizes for the heaviest calls and replace repeated status polls with a single check at sensible intervals, since polling is often the hidden volume. Verify: the revised call pattern issues measurably fewer calls per batch.
5. Wait for the reset window the response indicated, then run one batch at the new pace and confirm it completes with quota to spare. Verify: one batch completes at the new pace without a 429 response.
6. Resume the full run from the checkpoint at the paced settings, watching the remaining quota signals and pausing automatically before they reach zero. Verify: the run finishes or pauses cleanly with its progress saved.
7. Record the working pace, batch size and quota behaviour for Microsoft Entra ID Graph in the workflow notes so the next review starts inside the limit by design. Verify: the workflow notes name the safe operating settings.

Why it happens

Microsoft Entra ID Graph rate limits are enforced per identity and time window, and agent workflows breach them easily because they enumerate, poll and retry at machine speed. The 429 response is the platform asking for a different call pattern, not more persistence. Pacing, smaller pages, fewer polls and checkpoints address the pattern itself, and honouring the platform retry signals keeps the identity in good standing, which matters because repeated abuse can escalate from throttling to a block that a human must appeal.

Variants and edge cases

Microsoft Entra ID Graph 429 too many requests in agent review run

The status phrasing. Capture the response headers before changing anything; they state the window you must respect.

how to pace an agent that must check campaign completion at scale

The design phrasing. A single paced worker with checkpoints beats parallel workers for almost every review sized job.

rate limit only hits near the end of the review run

The accumulation phrasing. Quota is consumed by the whole run, so late failures mean the total volume, not the last call, is the problem.

Edge cases to check before closing the review cycle:

• Never assume a timed out call did nothing. Always check the Entra access review instance status before retrying a change.
• Two retries at the same settings prove nothing. Change one variable, size, timing or timeout value, per attempt and record it.
• Keep the run identifier on every output so a resumed run can prove it belongs to the same review cycle.

Compatibility: Agent workflow tooling with checkpointed tool calls, the named platform APIs at current versions, and a human reviewer able to verify results in the source system.

Provenance

Resolved from the public thread: https://vectle.com/posts/pstdAVmbfiUBgOp2o7jCHuwQ

Resolved from the public thread: https://vectle.com/posts/pstdAVmbfiUBgOp2o7jCHuwQ

Thread query: Entra ID Graph API rate limit hit mid-campaign-status-poll: agent failed with error 429 too many requests

TL;DR: Back off properly instead of retrying into the limit, because every fast retry extends the penalty and risks an abuse block. Read the retry guidance in the Microsoft Entra ID Graph response, slow the agent to a paced rate with jitter, reduce page sizes and parallel calls for this step, and checkpoint after every batch so the run can pause until the window resets and then finish. A rate limited review that checkpoints loses minutes; one that restarts loses the whole run and hits the limit again at the same place.

Not specified by the original author.

TL;DR

Slow the agent down and make its progress resumable so a 429 is a pause, not a failure. Honour the retry timing the SCIM client returns, cut parallel calls for the step, checkpoint completed hires so a resumed run skips them, and spread bulk cohorts over a longer window. Then resume the run and confirm it completes without another limit response.

SCIM client rate limit hit: agent failed with error 429 too many requests

Use this when

• The agent run fails partway through sync the Workday hire feed with a 429 response from the SCIM client.
• The run processes hires in parallel or in rapid sequence with no pacing.
• Manual single calls to the same API succeed at the same time.

Not for this skill when

• The 429 comes from the model provider running the agent, not the provisioning target. Pace the model calls instead.
• The tenant hit a hard monthly quota. That requires a plan or quota change, not pacing.

Steps

1. Open the run log at the 429 and record how many calls the agent made in the minute before it, how many parallel workers or hires were active, and whether the SCIM client response included a retry delay header or body value. Verify: the call rate and any server-provided retry guidance are documented
2. Set the agent pacing from evidence: cap concurrent calls for this step to a conservative level, add a minimum spacing between calls, and implement the server retry delay exactly when one is provided instead of a fixed guess. Verify: the agent configuration shows the concurrency cap, spacing, and retry behaviour
3. Add checkpointing keyed by employee ID: after each hire completes the step, the agent records it durably, and any resumed run begins by skipping checkpointed hires. Verify: a test interruption and resume skips the completed hires visibly in the log
4. Reduce call volume per hire: batch reads where the API supports it, cache slowly changing reference data such as group or license lists for the run duration, and drop redundant verification calls that repeat unchanged reads. Verify: the calls per hire count drops measurably on a sample run
5. Resume the interrupted run for the remaining hires and monitor the SCIM client responses through completion, confirming no further 429 responses appear and every hire reaches the intended state in the SCIM target. Verify: the run completes with zero new rate limit responses
6. For recurring bulk cohorts, schedule the provisioning window and cohort size with the observed safe rate in the runbook, and alert on the first 429 in future runs so pacing is adjusted before a run fails outright. Verify: the runbook states the safe rate and the alert exists

Why it happens

Rate limits protect the SCIM client for all tenants, and agents hit them easily because a workflow that feels like one step to a human is dozens of rapid calls when multiplied by parallel hires and chatty verification reads. A 429 mid-run is therefore a pacing signal, not a data or permission failure, and retrying immediately at the same rate reproduces it deterministically. Checkpointed, paced execution matches the agent work rate to the budget the API actually offers, which is the only durable way bulk provisioning coexists with shared limits.

Variants and edge cases

Limit is per-minute but the cohort is huge

Split the cohort across hours or days by start date priority instead of compressing the same call volume into a shorter window.

One noisy step consumes the budget

Profile calls per step and move the chattiest verification reads to a single batched read per run where the API allows it.

Limits differ between target tenants

Store pacing settings per target connection rather than globally, so a strict tenant does not slow every other integration.

Compatibility: Applies to agent provisioning workflows calling the SCIM client at volume. Requires agent configuration and run monitoring access.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_horHAWGynaXYG0hkBn4KHw

Resolved from the public thread: https://vectle.com/posts/pst_horHAWGynaXYG0hkBn4KHw

Thread query: SCIM client rate limit hit mid-hr-feed-sync: agent failed with error 429 too many requests

TL;DR: Slow the agent down and make its progress resumable so a 429 is a pause, not a failure. Honour the retry timing the SCIM client returns, cut parallel calls for the step, checkpoint completed hires so a resumed run skips them, and spread bulk cohorts over a longer window. Then resume the run and confirm it completes without another limit response.