How long coding agents actually take

A speech bubble, dashed and fading, holds a wall calendar shedding blank pages. Beside it a solid stopwatch, its hand a terminal cursor, is measured by two small cursor robots with a tape measure.

Measured from 3,383 completed agent runs, not from anyone's opinion. Half finished within 2 min. One in ten took more than 20 min.

Ask an assistant how long a feature will take and it answers in human-team time. agentpace answers from measured agent runs instead: P50, P80 and P90 for your own plan of tasks, with the human part stated on its own line. Logging your own runs works with Claude Code subagents and Codex today.

Try it, with no install

Needs uv. The seed data and the example plan ship inside the package.

uvx --from git+https://github.com/mohamedzhioua/agentpace agentpace estimate --example

It estimates this plan, 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"] }
  ]
}

Every completed run from 2026-W40 to 2026-W45. Time runs left to right on a log scale.

Times by model, kind and risk

Completed runs only. Grey rows have fewer than 10 runs, so treat them as anecdotes.

Model Kind Risk

This is not a model speed ranking

People send harder work to stronger models, so a slower row may just mean harder tasks.

This is one person's workflow, so far

3,430 runs in 293 sessions from 1 contributor, with heavy automated fan-out. It will not match your setup until more people contribute.

How well the forecasts hold up

On 56 held-out sessions, the P80 for the whole session covered 80% and the P90 covered 91%. For a single task the median miss is 78%, so read a range, never one number.

Check it yourself

uv tool install git+https://github.com/mohamedzhioua/agentpace
agentpace ingest                    # your own runs, stored locally
agentpace estimate plan.json        # P50, P80 and P90 for a plan, like the one above
agentpace share --dry-run           # the exact rows you would contribute