Show HN: Agent-estimate, how long a coding task takes, at agent speed

I have used Codex & Claude Code for coding for a while, but how long a coding task will actually take? When I ask Claude Code to estimate, the result is often from training data, which is based on human speed. That’s why I built this tool, to estimate effort in ai agent speed. I run it every morning before I dispatch coding tasks to my agents.

What's in it: task sizing: auto-classifies XS to XL from the description, then runs PERT on that tier human-equivalent comparison: a per-task-type multiplier so you see the speedup METR p80 thresholds: warns when an estimate exceeds a model's reliability horizon wave planning: schedules independent tasks in parallel across a multi-agent fleet

The estimation data is from my daily coding tasks from past few weeks: per-runtime calibration: Opus 4.7, GPT-5.5, different models have different reliability horizons and costs per-task-type priors: backend, frontend, app development, docs, and brainstorm PR review: I usually let Codex and Claude Code review each other’s code, and the tool takes that into consideration a calibration loop that keeps me honest: dispatch data is validated at end of day by my coordinator agent

Try it: pip install agent-estimate, read the code https://github.com/kiloloop/agent-estimate/ , or the writeup https://kiloloop.com/agent-estimate/

Some background on this: I run 3-4 coding agents in parallel every day, every morning I ask myself, what realistic goal can I deliver during my 8h working time? Thus at the beginning of the day, I will always work with my coordinator agent, to pick up today’s task queue, estimate them and go on execution.

If you ask an AI agent how long a task will take, it answers in human time, so the gut number is that’s an afternoon when the real answer is 20min. I have run it for a few months now, with a few hundred real dispatches. With better estimation my morning plan and my end-of-day reality mostly line up now, and when they don't it's usually because something finished faster than I guessed, not slower.