agentpace

Stop estimating agent work in human weeks. agentpace answers "how long will this take?" from measured agent runs: a P50/P80/P90 range, with the human part (review, testing, decisions) stated on its own line and never derived from agent time.

Try it

With uv, this needs no install, no clone and no history of your own:

uvx agentpace estimate --example

<!-- claim:example_output -->
Plan total on the session clock: P50 2.4 h, P80 5.9 h, P90 9.0 h
The session clock counts agent runs and the gaps between them, each gap up to the idle cap (900 s by default), so waiting on you between runs sits inside the total up to that cap.
Human part (review, testing, decisions, waiting between agent turns): not measured. No allowance for it is added to the total above. State it separately; do not derive it from the agent time.

task          P50        P80        P90  critical  cell (n)
schema     30 min     85 min      2.5 h     100%  feature x normal x claude-opus-5-5 (n=205)
api        30 min     87 min      2.5 h      53%  feature x normal x claude-opus-5-5 (n=205)
ui         30 min     86 min      2.6 h      47%  feature x normal x claude-opus-5-5 (n=205)
docs         72 s      6 min     13 min      18%  docs x normal x claude-opus-5-5 (n=0)
review      3 min      8 min     13 min      82%  review x low x claude-opus-5-5 (n=1255)

Critical path: schema -> ui -> review (longest in 47% of simulations)
Effective parallelism at cap 3: 0.90 (from 29 sessions). Throughput set the total in 76% of simulations.
Method: shrunk. 3383 runs, 47 not completed so excluded. 20000 simulations, seed 20261107.
Not completed, by harness: claude-code 13 of 3219 (0.4%; 3 failed, 10 unknown), codex 34 of 211 (16.1%; 33 failed, 1 unknown). These are recorded harness outcomes, not task acceptance, and they are left out of the fit.

Warnings:
  - cell docs x normal x claude-opus-5-5: n=0. No recorded run has kind docs at risk normal with model claude-opus-5-5. Mean from docs x normal (n=21); residuals pooled from all; interval widened by the spread between models. Recorded models under docs x normal: claude-sonnet-5-5 (21).
<!-- /claim -->

The plan it estimated is examples/plan.json, where each task is one agent run:

{
  "agent_cap": 3,
  "model": "claude-opus-5-5",
  "tasks": [
    { "id": "schema", "kind": "feature", "risk": "normal" },
    { "id": "api", "kind": "feature", "risk": "normal", "depends_on": ["schema"] },
    { "id": "ui", "kind": "feature", "risk": "normal", "depends_on": ["schema"] },
    { "id": "docs", "kind": "docs", "risk": "normal", "depends_on": ["api"] },
    { "id": "review", "kind": "review", "risk": "low", "depends_on": ["api", "ui"] }
  ]
}

To estimate your own, write a plan like it and run agentpace estimate plan.json, or pass - to read it from stdin. In a plan, depends_on and model are optional, and risk can be left out: it defaults to the risk the rows record for the task's kind, because in the data risk is the agent's role, not a rating of the work.

Reading the output

• The plan total is a range, not one number, because agent run times have a long tail and any single number misses most of the time. P50 is a coin flip. P80 is the forecast's 80th percentile, the one to plan on, not a promise: the backtest shows how often it held.
• The session clock counts agent runs and the gaps between them, each gap up to the idle cap, so a short wait on you sits inside the total and a longer one is clipped at the cap.
• The human part (review, testing, decisions, waiting between agent turns) is not measured, and no allowance for it is added. It stays on its own line, and agentpace never derives it from agent time.
• cell (n) is the group of recorded runs a task falls in, by kind, risk and model, and how many runs it holds. critical is the share of simulations in which the task was on the longest chain of dependent tasks.
• Effective parallelism is work divided by elapsed time, learned from recorded sessions at the plan's agent cap. It can be below 1, because a session's clock includes the time between runs.
• Warnings are part of the answer. When a task's cell has no or few recorded runs, one says so, what the range could not use (the kind, the risk level or the model) and where it leaned instead.

In your agent

npx skills add mohamedzhioua/agentpace

or, in Claude Code 2.1.275 or later, as a plugin:

/plugin install agentpace --marketplace mohamedzhioua/agentpace

Either way, install the CLI first (Install). npx skills add gives /pace-estimate, /pace-log and /pace-share; the plugin namespaces them as /agentpace:pace-estimate, /agentpace:pace-log and /agentpace:pace-share. The agentpace skill triggers by itself in both. Install one route, not both, or each skill loads twice.

Then ask your agent "How long will the billing feature take?": the skill calls the CLI.

Why agentpace

Ask an assistant how long a feature will take and it answers in human-team time, "a few weeks", even when agents will do the work. agentpace answers from measured agent runs instead, as a standard-library Python CLI, a set of agent skills and an open CC0 dataset, and it publishes how often it is wrong.

These are 3,383 completed agent runs: Claude Code subagent runs and Codex turns. Half finished within 2 min, and one in ten took more than 20 min. All of them come from one person so far, so read Limitations before trusting any of this, and consider contributing your runs: issue #2 is the standing call.

Does it fit your setup?

Install

uv tool install agentpace

pipx install agentpace works too. It needs Python 3.10 or later and has no dependencies. After installing, drop uvx from the command above.

For the agent skills, use In your agent: the npx command, or the plugin in Claude Code. Without Node, copy the four folders under skills/ into your agent's skills folder, for example ~/.claude/skills/ for Claude Code. No skill-capable agent? Paste the portable prompt into any assistant.

Skills

agentpace triggers by itself on questions about how long software work will take. The three /pace-* skills run when you call them. The portable prompt gives a rule-of-thumb version of the same answer in assistants without skills.

Re-answer a quoted timeline

Suppose someone has called a customer-import feature "a 3-month project". The skill splits it into agent tasks, as in examples/quote-plan.json, and passes the quote on exactly as it was given:

agentpace estimate examples/quote-plan.json --quoted "3 months"

<!-- claim:quote_example_output -->
Plan total on the session clock: P50 4.3 h, P80 10.3 h, P90 15.7 h
Quoted: "3 months" (calendar time, as quoted; unverified, not converted)
The session clock counts agent runs and the gaps between them, each gap up to the idle cap (900 s by default), so waiting on you between runs sits inside the total up to that cap.
Human part (review, testing, decisions, waiting between agent turns): not measured. No allowance for it is added to the total above. State it separately; do not derive it from the agent time.

task               P50        P80        P90  critical  cell (n)
schema          30 min     85 min      2.5 h     100%  feature x normal x claude-opus-5-5 (n=205)
importer        30 min     86 min      2.6 h      40%  feature x normal x claude-opus-5-5 (n=205)
api             30 min     87 min      2.5 h      60%  feature x normal x claude-opus-5-5 (n=205)
settings-ui     31 min     86 min      2.5 h      32%  feature x normal x claude-opus-5-5 (n=205)
tests           38 min      2.1 h      3.8 h      68%  tests x normal x claude-opus-5-5 (n=2)
review           3 min      8 min     13 min     100%  review x low x claude-opus-5-5 (n=1255)

Critical path: schema -> importer -> tests -> review (longest in 40% of simulations)
Effective parallelism at cap 3: 0.90 (from 29 sessions). Throughput set the total in 78% of simulations.
Method: shrunk. 3383 runs, 47 not completed so excluded. 20000 simulations, seed 20261107.
Not completed, by harness: claude-code 13 of 3219 (0.4%; 3 failed, 10 unknown), codex 34 of 211 (16.1%; 33 failed, 1 unknown). These are recorded harness outcomes, not task acceptance, and they are left out of the fit.

Warnings:
  - cell tests x normal x claude-opus-5-5: n=2, effective n=2.0, below 30. Mean shrunk toward broader cells; residuals pooled from all; interval widened.
<!-- /claim -->

The quote is calendar time, as quoted, and nothing has checked it. The range is built from measured agent time on the session clock, which counts the gaps between runs up to the idle cap. The human part is stated apart, on its own line, and no allowance for it is added to either. The quote is shown beside the range, never converted into agent time or set against it.

Commands

agentpace ingest reads Claude Code subagent transcripts from ~/.claude/projects and keeps rows in ~/.agentpace, on your machine. A second run reads only what changed. To log every run as it finishes, add the opt-in hook that /pace-log shows.

There is also a leaderboard of median, P75 and P90 by model, kind and risk.

Backtest: how often it is wrong

Bottom line. On 56 held-out sessions, the whole-session P80 covered 80% (CI 70% to 91%) and the P90 covered 91% (CI 84% to 98%). For a single task, the P80 covered 77.7% (CI 71.4% to 83.8%), but the median miss of its P50 is 78%, so plan on the range, never the middle. Calibration drifts from week to week: per-session P80 coverage ran from 66% to 86%.

Every number here comes from the published rows. agentpace backtest regenerates them, and CI fails if they drift. The headline scores Claude Code sessions, the population the method was chosen on. Codex is scored on its own. The full report has every split, every baseline and every loss.

Forecasts use only what is known before a run: model, kind and risk. Whole sessions are held out, so no run can leak into its own forecast. These headline numbers come from the same time folds that chose the method, so treat them as slightly optimistic. Weeks after 2026-W45 form a locked holdout that never takes part in selection. Its status: not enough data (no weeks after 2026-W45 yet). Every interval below is a 95% bootstrap interval that resamples whole sessions.

Where it falls short.

• The median error on a single task is 78%. Use the range, not the middle.
• Pooled, P80 coverage is 88.8% when each run counts equally rather than each session, so the ranges are too wide for the small, fast runs that dominate the row count.
• A global median beats agentpace on MAPE (234% against 494%), and a sum of medians beats it on plan MAPE (85% against 249%), though that sum covers only 5% of sessions at P80. MAPE punishes the long tail of over-forecasts that honest ranges need.
• The report lists 21 places where a baseline won. That list is generated, so it cannot be edited out.
• On the project split, which runs over both harnesses and where the largest project, a codex one, holds 137 of 212 held-out sessions, P80 coverage is 88.9%. The contributor split could not run (1 contributor).
• The whole-session numbers above hold out later weeks, not codebases. With whole projects held out instead (63 sessions across 11 projects, each predicted from every run outside its project, later weeks included), the P80 covered 87% (CI 79% to 95%) and the P90 95% (CI 89% to 100%), and the 95% interval for a sum of medians' pinball loss minus agentpace's is -7019 to +27430; only an interval wholly above zero shows that baseline to be worse. It is a retrospective check on 1 contributor's projects, and it never chooses the method. The report has each project on its own.
• The session-weighted design this project started with did not beat the simpler unweighted one, so the simpler one shipped. The selection rule was fixed before the first result. Scoring it on Claude Code sessions only was decided after a run with Codex rows pooled in had been seen, and that pooled result is printed in the report.

Privacy

agentpace never stores prompts, code or commit messages, and the rows you publish hold no content, file paths or repo names. Paths live in one local place, the ingest cache in ~/.agentpace/, which is never shared. It holds two kinds: the path of each transcript, so the next ingest re-reads only what changed, and the project folder (for Codex, the working folder), which reaches a row only as a keyed hash.

• What gets read. From each transcript: timestamps, model ids, token counts, tool names, error flags and two metadata keys, plus the project folder as a hash input. JSON is decoded one line at a time, so other text passes through memory, but only those values are copied out. A test plants marker strings in every sensitive field and proves none reach any output.
• What a row holds. Only the fields in schema/run.schema.json. Every string is an enum or a strict pattern, so free text has nowhere to land.
• Who you are. Keyed hashes under a random secret that never leaves your machine. No email, handle or hostname goes in, so nobody can reverse them by guessing names.
• Nothing is sent. There is no network code in this package at all, and a test fails if any is added. Contributing is a git push you run yourself. The site loads nothing from anywhere else.

Rows are not anonymous in every sense. They link to each other, they show roughly when you worked, and a pull request ties them to your GitHub account. agentpace share says so before it writes anything.

Contribute your runs

Sharing is opt-in, and your rows join a public-domain dataset.

uv tool install agentpace    # or pipx install agentpace
agentpace ingest
agentpace share --dry-run    # prints the exact rows and writes nothing

Read the rows. If you are comfortable with them, clone your fork and write the file into it:

git clone https://github.com/<your-fork>/agentpace && cd agentpace
agentpace share --write      # adds or extends one file under data/runs, and still sends nothing
agentpace validate data/runs # the same checks CI runs

Then open a pull request with that one file. A data pull request dedicates your rows to the public domain under CC0-1.0. See CONTRIBUTING.md.

The most useful thing you can do is contribute: the contributor split cannot run until a second person does.

How it compares

Audited at the commits listed in docs/DESIGN.md.

Each of these does something better than agentpace:

• agent-estimate binds forecasts to receipts, so one issued after work started cannot score. It also forecasts token cost, and ships as a pip package and a GitHub Action.
• claude-eta gives live per-turn ETAs inside a session, evaluates itself walk-forward on your own history, and has consent-gated community sharing and a repair-loop detector.
• agent-estimation is the simplest and most teachable model here, and it needs no data at all.
• web-dev-estimation has the best skill structure, with batch and re-estimate modes that work from documents before any history exists.
• mythical-agent-month separates machine, human and waiting time, tracks usage budgets, and openly publishes its own held-out miss.

Limitations

• One person's workflow. Every run comes from 1 contributor, across 21 projects and 293 sessions, from 2026-W40 to 2026-W45. One project holds 67% of runs, and three sessions hold 45%. It includes heavy automated fan-out, so it may not resemble your setup.
• Not a model speed ranking. People send harder work to stronger models, so a slower row may just mean harder tasks.
• feature means any implementation work. Without reading prompts, metadata cannot separate features, bug fixes and refactors. Those kinds stay thin until people label runs.
• Retries are not separated. A run's duration includes any test-fix loop inside it. retries counts errored tool results, a proxy, because a real retry cannot be seen without reading content, and no field that describes what happened in a run (retries, tool_calls, tokens, outcome) ever predicts a duration. Codex upkeep sessions (compact) are skipped; nothing else is split out.
• Pooled by kind, risk and model only. A cell with fewer than 30 runs leans on broader cells, with a warning. Repository size is not recorded, and size is an edit count measured after the run, so neither can segment the published runs. Long integration work has no cell of its own. risk in the rows is the agent's role (review and research agents low, migration agents high, all else normal), and a plan's risk is read the same way, so marking a task high does not lengthen its range.
• Two harnesses, still one person. The published rows hold Claude Code subagent runs and, since 0.6.0, Codex turns, all from the same contributor. Codex is scored on its own, never in the headline. Each Codex exec session is a single turn, so the session effect and the parallelism factor learn nothing from Codex, and a Codex plan borrows Claude Code's session structure. Gemini CLI, OpenCode and Cline have experimental adapters: built from each tool's source, not yet verified on real sessions, with no published rows (validate rejects them until each adapter is verified; how to verify one), and problems go in an issue. Other harnesses have no adapter.
• No human allowance, and short waits sit inside the clock. The plan total is a forecast on the session clock, which counts agent runs and the gaps between them, each gap up to the idle cap (900 s in the published rows). So a wait on you shorter than that sits inside the total, a longer one is clipped at the cap, and no allowance for review, testing or decisions is added. Every estimate says so on its own lines and never derives the human part from agent time. agentpace turns measures your own wait per human turn from your main-session timestamps and keeps it on your machine. That is elapsed time, not effort, shown beside the range and never added to it.
• Failed runs are under-counted. Only runs with a clear completion signal train the estimator, so P90 may run low. Recorded runs that did not complete (failed or unknown), by harness: claude-code 13 of 3,219 (0.4%), codex 34 of 211 (16.1%). Those are harness outcomes, not task acceptance, and every estimate prints the counts for the rows it loaded.
• Size is not used by default. Stored size is measured after a run, so it would leak the answer into the forecast.
• The transcript format is undocumented. It may change, and Claude Code removes old transcripts, so run agentpace ingest regularly.
• Dependencies are untested. The whole-session test cannot see real dependencies.

Development

PYTHONPATH=src python -m unittest discover -s tests -t .

The tests need nothing installed. All the math is in src/agentpace/estimate.py. The reasoning behind the design is in docs/DESIGN.md.

License

Code: MIT. Data: CC0-1.0.