More screenshots
Works with the AI Agents you already use
Agent Timeout Budget — Where Your Latency Goes
Reads run traces and computes per-tool p50/p95/p99 latency, timeout rate, retry amplification and a suggested timeout from the observed tail.
Free
Agent Timeout Budget — Where Your Latency Goes
Example session with this skill installed
python3 scripts/timeout_budget.py traces.jsonl --fail-on high
{"run_id":"run-005","step":1,"tool":"search","duration_ms":45000,
"status":"error","error":"timeout","retry":true,"side_effect":false}
- Read your context and instructions
- Compiled the agent timeout budget
agent-timeout-budget 40 runs 320 steps 5 tools
run p50 30.1s p95 56.7s p99 220.6s
retries 32 amplification 1.11x steps p95 10
TOOL CALLS P50 P95 P99 T/O RETRY SET
search 72 7.4s 45.0s 45.0s 8 18 67.5s
send_email 56 1.5s 2.0s 2.0s 0 14 3.5s
HIGH B1 tail latency search p99 45.0s, max 45.0s
HIGH B2 timeout rate search 8/72 calls timed out (11.1%)
HIGH B4 retried side effect send_email 14 retries
HIGH B5 run budget run p99 run duration 221s
FAIL findings at or above 'high'
Connects securely to your tools. The creator never sees your data.
About this skill
A timeout set from the median fails the tail. A retry added without a budget multiplies attempts across layers. This reads the traces and computes the numbers a timeout policy needs, per tool and per run.
It reports p50, p95 and p99 latency for every tool, the timeout rate, the number of retries, and retry amplification across the run. It flags side-effecting tools that are being retried, because a duplicated send or payment is worse than a failed one. It computes run-level p99 and p95 against an end-to-end budget, and suggests a timeout at 1.5× the observed p99, rounded up to 500ms.
- Per-tool tail latency, including max, from the trace you already have
- Timeout rate against the tool's own tail, not a universal default
- Retry amplification across all attempts, so independent retry layers are visible
- Retried side effects, with the tool and count
- Run p99 and step-count pressure, so a fast tool cannot hide a slow run
- Exit 1 at a severity threshold for CI
The shipped failing trace is 40 runs and 320 steps: search p99 at 45s, 8 timeouts in 72 calls, 14 retries on a send tool, and a run p99 of 221s. The clean trace is 28 runs with no findings.
Good to know: quantiles from small samples are unstable, and timeouts censor the tail, so the slowest calls may be missing because they were cut off. A trace cannot prove a remote operation stopped after the client timed out, or that an idempotency key prevented a duplicate. The side-effect check is name-based and conservative; confirm the flag in your own registry. Suggested timeouts are a starting point, not a policy: the real cap is the lesser of the step budget and the time left before the run deadline.
Small sample sizes produce unstable quantiles. Timeouts in traces can censor the true tail. Side-effect detection requires manual confirmation of flags in your tool registry.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Free forever
No account required to browse
Trust & safety
Security scan
Security scan passed 1 day ago
- Free with an Agensi account