Costs and numbers that change the decision
At scale, the cost difference between these two thresholds is substantial. Consider a scenario with 1,000 runs and a 1% stuck rate, yielding 10 stuck agents. A 15-minute terminal timeout spends 10 multiplied by 15, or 150 agent-minutes on these stuck runs. In contrast, a 3-minute kill spends 10 multiplied by 3, or 30 agent-minutes. The delta between the two strategies is 120 agent-minutes reclaimed per 1,000-run window. This arithmetic converts the 3/15 split into heartbeat-count and agent-minute calculations, making the case for the tighter probe budget the default choice for cost-conscious fleets.
Grounding these numbers in the 2026 public record, Gamma demonstrates that AI generation completes in 1-2 minutes while total creation, including review, spans 15-30 minutes; these are separate clocks, so a single 3/15 threshold cannot govern all agent fleets. Artificial Analysis Decision Index v2 reports cost per task ranging from $14 per 1M Tasks for Celeris-1.1 Decision to $859 for Clef, illustrating that resource consumption varies by model and task type. These figures confirm that the 3-minute probe with heartbeat-gated 15-minute timeout is not a universal constant but a calibrated default that must be adjusted to the specific cost profile of the agent fleet in operation.
Limits: where the timeout rule breaks
| Edge Case | Wait / Condition | Rule Break | Rule Wins |
|---|---|---|---|
| External batch tool | Wait longer than 3‑minute health‑check budget | Breaks at 3 min | Wins at 15 min only with ≤60 s heartbeat and idempotent retry |
| Human approval (sync) | Synchronous step after health‑check | Breaks both 3 min and 15 min timeouts | Resolved by moving to async state machine |
| No heartbeat / no persisted state | Missing heartbeat or state persistence | 15 min timeout fails (stall indistinguishable) | 3 min terminal kill remains effective |
Human approval performed synchronously creates a bottleneck that can defeat both the 3-minute and 15-minute timeout rules; moving the approval step into an asynchronous state machine removes it from the health-check path. Before citing any outside roll-call training example, verify that the source actually measures synchronous interruptions in your context; the cited Police1 item is about roll-call training, not AI agent health checks.
These edge cases show where the timeout rule breaks: external batch tools may need longer than the 3-minute health-check budget; synchronous human approval can block both the 3-minute and 15-minute rules; and a run without heartbeat or persisted state cannot qualify for the extended timeout. The three-control design—heartbeat, boundary-state persistence, and idempotent retry—keeps the fleet enforcing its health-check budget under those conditions.
And copyable health-check sheet
The illustrative fleet of 100 runs establishes the operational envelope for the 3-minute p95 rule. Three agents process the workload, with an 180-second probe budget and a 900-second terminal timeout governing the overall lifecycle. Of the 100 runs, 72 finish under 3 minutes, 8 batch runs consume the full 8-minute tool duration, and 20 runs stall after 6 minutes with no heartbeat signal. This distribution creates the baseline from which the tiered health-check mechanism derives its thresholds.
Checkpoints execute at fixed intervals to enforce the three-tier probe/watch/terminate logic. At the 180-second mark, 72 runs pass and 28 enter the watch window. By 360 seconds, the 8 batch runs transmit a heartbeat and 20 stalled runs have already missed 12 of the expected 30-second beats. At 540 seconds, the batch runs continue their heartbeat stream, while the stalled runs remain silent. These checkpoints are not arbitrary; they are the decision points that trigger the terminal timeout only when the heartbeat gate is satisfied.
The terminal timeout of 15 minutes applies exclusively to runs that emit a heartbeat at least every 60 seconds and persist state at every tool boundary. In the illustrative window, the 8 batch runs meet these conditions and are granted the extended 13 minutes 10 seconds before eventual return. The 20 stalled runs, lacking both a heartbeat and persisted state, are terminated well before the 900-second cap, freeing resources for the 72 completing runs. This distinction proves that the 3/15 split is not a fixed constant but a conditional gate dependent on runtime signals.
Decision rules for your agent fleet
The decision logic relies on concrete measurements taken over 100-run windows. Operators should monitor health-check p95 across these windows; if p95 exceeds 3 minutes in two consecutive windows, alert and investigate latency, and do not grant the 15-minute terminal timeout unless the run meets all three controls. If a run satisfies heartbeat at or under 60 seconds, state persisted at every tool boundary, and idempotent retry, the 15-minute timeout becomes an available option for that qualifying run. Evaluate the controls per run; the article does not establish a fleet-wide uniformity rule beyond those measurements.
What to do next
| Step | Action | Why it matters |
|---|---|---|
| 1 | Baseline the fleet: measure health-check p95 across 100-run windows and record the current value against the 3-minute default. | The 3-minute p95 budget is only enforceable against a measured baseline; the 100-run window is the unit of every downstream decision. |
| 2 | Configure the alert to fire only when p95 exceeds 3 minutes in two consecutive 100-run windows. | Two consecutive windows separate sustained degradation from a one-off spike, preventing false pages. |
| 3 | Verify each agent emits a heartbeat at least every 60 seconds before granting the 15-minute terminal timeout. | Heartbeat interval at or under 60 seconds is the first of the three mandatory controls. |
| 4 | Persist state at every tool boundary and audit the writes in staging. | Per-boundary persistence is required alongside heartbeat and idempotent retry; a gap here disqualifies the extended timeout. |
| 5 | Make every tool retry idempotent and prove it with a forced double-execution test. | Without idempotent retry, the 15-minute terminal timeout does not apply — full stop. |
| 6 | If any of the three controls is false, hard-kill the agent at 3 minutes instead of waiting for the 15-minute timeout. | All three controls must be present; otherwise the extended timeout risks unsafe retries on corrupted state. |
Frequently Asked Questions
What is the default timeout threshold for the 3-minute probe?
The 3-minute probe with heartbeat-gated 15-minute timeout is the calibrated default.
How many agent-minutes are reclaimed per 1,000-run window by choosing a 3-minute kill over a 15-minute timeout?
120 agent-minutes reclaimed per 1,000-run window.
What is the cost per task for Celeris-1.1 Decision according to Artificial Analysis Decision Index v2?
$14 per 1M Tasks for Celeris-1.1 Decision.
What is the cost per task for Clef according to Artificial Analysis Decision Index v2?
$859 per 1M Tasks for Clef.
What is the AI generation time range that grounds the 2026 public record figures?
AI generation completes in 1-2 minutes.
What is the total creation time range including review that grounds the 2026 public record figures?
total creation, including review, spans 15-30 minutes
Quick answers
| What is the agent-minute cost difference between a 15-minute timeout and a 3-minute kill for 1,000 runs with a 1% stuck rate? | 120 agent-minutes reclaimed per 1,000-run window |
| How does the 3-minute latency budget relate to the 15-minute timeout in the context of AI generation and review clocks? | AI generation completes in 1-2 minutes while total creation, including review, spans 15-30 minutes; these are separate clocks, so a single 3/15 threshold cannot govern all agent fleets. |
| What does the Artificial Analysis Decision Index v2 report about cost per task for different models? | cost per task ranging from $14 per 1M Tasks for Celeris-1.1 Decision to $859 for Clef |
| Why is the 3-minute probe with heartbeat-gated 15-minute timeout considered a calibrated default rather than a universal constant? | resource consumption varies by model and task type, confirming that the threshold must be adjusted to the specific cost profile of the agent fleet in operation. |
| What is the primary reason for choosing a 3-minute latency budget over a 15-minute timeout in cost-conscious AI agent fleets? | The 3-minute kill strategy reclaims 120 agent-minutes per 1,000-run window compared to a 15-minute timeout, making it the default choice for cost efficiency. |
Also worth reading: Audit and trace AI agent decision chains: Audit and trace AI agent · Human-in-the-loop agent workflows: 7 best practices that scale: Human-in-the-loop agent workflows: 7 best · Human-in-the-loop approvals for critical AI agent decisions: Human-in-the-loop approvals for critical AI