$ agent-estimate --compare

Three ways
to know.

Q: How accurate is agent-estimate?

agent-estimate validate scores each forecast against actual work or wall minutes on the matching basis, so you measure accuracy on your own record rather than trusting a headline share. The maintainers publish their own estimate-vs-actual record as "Estimate accuracy, measured" on kiloloop.com/research.

No estimation, manual, or agent-estimate — three approaches to scoping AI-agent work. Here's how they actually compare, on the dimensions that matter.

· 250+ dispatches validated (May 2026)· 7 dimensions· no marketing

agent-estimate --compare --tldrzsh

$ agent-estimate --compare --tldr
 
scoring against: speed, accuracy, AI-fit, ci, ux
────────────────────────────────────────────────
 
  ✗ --no-estimate    0/5  risky for anything with a deadline
  ~ --manual         2/5  slow, optimistic, AI-blind
  ✓ --agent-estimate 5/5  calibrated, fast, CI-ready
 
recommendation ─────────────────────────────────
  use agent-estimate when AI agents touch your sprint.
  use manual for judgment-heavy planning.
  use no estimate for throwaway code only.
 
  → see full breakdown below ↓

OVERVIEW

At a glance.

--no-estimateRISKY

No estimation

// Ship and hope.

Zero upfront cost, but scope creep, missed deadlines, and surprise failures. Fine for prototypes and throwaway experiments. Dangerous for anything with a deadline.

--manualSLOW

Manual estimation

// Human guesses in planning meetings.

Slow (30–60 min per sprint), systematically optimistic, and blind to AI-specific speed differences. The estimates degrade as model capabilities change quarter to quarter.

--agent-estimateRECOMMENDED

agent-estimate

// PERT math + METR data + dependency analysis.

Calibrated three-point estimates in ~2 seconds. Accounts for model-specific reliability limits, review overhead, and multi-agent parallelization. Scores forecasts against actuals via agent-estimate validate.

DIFF

Side-by-side, dimension by dimension.

$ agent-estimate --compare --table

agent-estimate compare · tablediff

dimension--no-est--manual--agent-estimate

Time to estimate-0~30–60 min/sprint+~2 seconds

AI speed awareness-no~rarely+per-model multipliers

Dependency analysis-no~whiteboard+automated wave planning

METR reliability-no~no+built-in thresholds

CI integration-no~no+GitHub Action

Review overhead-ignored~sometimes+3 modes (std/complex/3-round)

Scored vs actuals-no~sometimes+agent-estimate validate

7 dimensions · 3 approaches+7 ~0 -0 winner: agent-estimate

DECIDE

When to use each.

decide.shbash

# pick by project context
case $context in
 
  prototype)
    estimator="--no-estimate"
    # throwaway, near-zero cost of being wrong
    ;;
 
  mature_team)
    estimator="--manual"
    # judgment-heavy, distrusts automation
    ;;
 
  ai_workflow)
    estimator="--agent-estimate"
    # AI agents, CI budgets, repeatable
    ;;
 
esac

prototype→ --no-estimate

Prototyping, hackathons, throwaway scripts where the cost of being wrong is near zero.

mature_team→ --manual

Established velocity, judgment-heavy work (design, strategy), or stakeholders who distrust automation.

ai_workflow→ --agent-estimate

Sprint planning with AI agents, CI pipelines with time budgets, multi-agent workflows, repeatable estimates across changing models.

FAQ

Frequently asked questions.

?How accurate is agent-estimate?

›agent-estimate validate scores each forecast against actual work or wall minutes on the matching basis, so you measure accuracy on your own record rather than trusting a headline share. The maintainers publish their own estimate-vs-actual record as "Estimate accuracy, measured" on kiloloop.com/research.

?Does it work with models other than Claude?

›Yes. Built-in METR thresholds cover Claude (Opus, Sonnet, Haiku), GPT-5.5/5.4, and Gemini 3.1 Pro. Custom thresholds for any model in your agent config YAML.

?How does PERT estimation work?

›PERT uses three time estimates per task — optimistic, most likely, pessimistic — to compute a weighted expected duration: (O + 4M + P) / 6. Produces a probability distribution rather than a single-point guess.

?Can I use it without the CLI?

›Yes. The Claude Code plugin adds /estimate as a slash command. The GitHub Action runs estimates in CI and posts results as PR comments. Same structured output across all three.

$ start estimating

Stop guessing. Start shipping.

One command. No account. Reports in seconds.

$ pip install agent-estimate ← back to landing

Three waysto know.