$ agent-estimate --help

Know before
you build.

PERT estimates for AI-agent tasks. How long, which model is reliable enough, and what's the human-equivalent cost — in two seconds.

$ pip install agent-estimate

· apache 2.0 · python ≥ 3.10 · 250+ validated (May 2026)

agent-estimate · live output zsh

$ agent-estimate estimate --file tasks.txt
 
Task             Tier   PERT (O/M/P)    Expected   Human-eq
────────────────────────────────────────────────────────────────
Implement auth   M      25/50/90m       57.8m      160m
Add tests        S      12/23/40m       24.0m       75m
 
Timeline ────────────────────────────────
  best 37m   ·   expected 81.8m   ·   worst 130m
  human-equivalent: 235m  →  2.87× compression
 
  ⚠ METR warning: "Implement auth" exceeds Opus p80

PROBLEM

The estimation gap.

AI agents can write code — but how long will a task actually take? Manual estimation is slow and biased toward optimism. No estimation means scope creep and missed deadlines. The gap between "agents can do it" and "we know when it'll be done" is where projects break down.

agent-estimate closes this gap with PERT math, METR reliability data, and dependency-aware wave planning — calibrated forecasts in seconds, not meetings.

SYNOPSIS

Three steps. No meetings.

describe

Feed tasks via CLI args, GitHub issue refs, or a plain-text file. Markdown comments allowed.


                  # describe.txt
#42 — add OAuth flow
#43 — wire login UI
#44 — write E2E tests

estimate

Three-point PERT with spec-clarity, context, and agent-fit modifiers. Calibrated against 250+ dispatches (May 2026).


                  PERT = (O + 4M + P) / 6
25 + 4×50 + 90 = 52.5m
× 1.10 context tax
→ effective 57.8m

decide

Best / expected / worst, METR warnings, and per-agent load. Output as JSON, MD, or PR comment.


                  # timeline
best     70m
expected 90m
worst    130m
2.78× vs human

EXAMPLE OUTPUT

Wave plan: parallel across agents.

agent-estimate plan --waves json

# wave 1 — parallel execution (T+0 → T+58m)
┌────────── codex ──────────┐  ┌────────── claude ─────────┐
│ Implement auth            │  │ Add tests                 │
│ 57.8m + 17.5m review      │  │ 24.0m + 7.5m review       │
│ ! exceeds threshold       │  │ + within threshold        │
└───────────────────────────┘  └───────────────────────────┘
 
# wave 2 — sequential (T+58 → T+90m)
┌────────── gemini ─────────┐
│ Integration + ship        │
│ 32.0m + 8m review         │
└───────────────────────────┘
 
total fleet time: 90m  ·  human-eq: 250m  ·  2.78×

OPTIONS

What you get.

--pert

PERT estimation

Three-point optimistic/likely/pessimistic per task. No hand-waving, just math.

--metr

METR thresholds

Auto-warns when tasks exceed model reliability limits. Opus 90m, Sonnet 30m, GPT-5.5 90m.

--waves

Wave planning

Dependency-aware parallelization. See which tasks run concurrently across agents.

--human-eq

Human-equivalent ratios

AI speed comparison by task type: boilerplate 3–5×, feature 2–4×, bug fix 1.5–3×.

--multi-agent

Multi-agent sessions

Estimate across parallel agent sessions with per-agent load and cost breakdown.

--review

Review overhead

Three modes — standard, complex, 3-round. Factor in human review before it surprises you.

250+

dispatches validated · May 2026

3.7×

avg compression

75%

within our calibration band

AI runtimes tested

// production data — claude · codex · gemini · cline. not synthetic benchmarks.

$ start estimating

One command. No account.

Reports in seconds.

$ pip install agent-estimate Read the README →

Know before you build.