More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    Agent Timeout Budget — Where Your Latency Goes

    1

    Reads run traces and computes per-tool p50/p95/p99 latency, timeout rate, retry amplification and a suggested timeout from the observed tail.

    Free

    Security scanned
    Agent Timeout Budget — Where Your Latency Goes

    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.

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for LLM Ops.

    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. 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. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before buying a skill.

    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 to use

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 1 day ago

    • Free with an Agensi account

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Listed1 day ago

    What's inside

    Frequently Asked Questions

    The skill requires a JSONL trace file containing `run_id`, `tool`, and `duration_ms`. Because it processes standardized trace data, it is compatible with any AI agent framework (like LangChain, CrewAI, or AutoGPT) as long as you can export your execution logs to that format.

    By calculating p99 tail latencies, the skill helps you set timeouts that account for the slowest successful runs rather than the average. This prevents 'false-positive' failures where an agent is killed simply because a tool was slow, while also identifying where retries might trigger dangerous duplicate state changes.

    The skill is a standalone Python 3.10+ CLI tool that uses only the standard library, meaning it has zero external dependencies. You simply download the script, provide your JSONL trace file, and it generates a Markdown report or JSON summary locally on your machine.

    While you can analyze any amount of data, the tool includes a warning for small sample sizes where quantiles may be unstable. For the most reliable p99 budget suggestions, we recommend providing a trace file with at least 100-200 samples per tool.

    Updates are delivered through the marketplace's versioning system. Buyers receive lifetime access to the current version and all subsequent maintenance updates, including any future optimizations for larger trace file processing or additional report formats.

    New in LLM Ops

    Free