jull@monkeykode:~/calculator

Monkey Kode

Estimating work done by people who direct AI agents

MODEL
INPUTS
OUTPUT
SITE
calculator.monkeykode.com

Agents made some software work dramatically faster and left other work exactly as slow as it was. This formula takes a conventional estimate and adjusts it for a team where each person directs a handful of agents, so the time and cost I quote reflect how the work actually gets done.

The formula

Human hours equal H times the sum of one minus f, plus f divided by s times C of a, all multiplied by one plus gamma times P minus one. C of a equals a divided by one plus alpha times a minus one, plus beta times a times a minus one. Price equals hours times rate times one plus buffer. Duration equals hours divided by P times focused hours per day.

Point at any symbol to see which input it stands for in the calculator below.

Try it with your project

The project and the team

What an experienced team would quote without agents.
Tune the assumptions
Lower it for integrations with older systems or projects with a lot of discovery.
Hours of conventional work delivered per hour a person spends directing one agent.

The estimate

Human hours40.110.0× faster than conventional
Price$4,606at $100 an hour, 15% buffer
Calendar time6.7 daysabout 1.3 working weeks
Conventional
400 h
With agents
40.1 h
400 × [0.05 + 0.95 ÷ (9 × 2.11)] × 1.00 = 40.1 hours
At the sweet spotEffective agents per person 2.11 of 4Best count 4Ceiling 20×
Speedup by agents per person

How it works

  1. Start with the conventional estimate

    H is the number of hours an experienced team would need without agents. It is still the best measure of how much work a project contains, so everything else adjusts it.

  2. Split it into two kinds of work

    A share f is work agents can carry out, such as writing code, tests, migrations, and documentation. The remainder is work only people can do: understanding what you need, making decisions, and signing off on the result. That part takes as long as it always did.

  3. Speed up the agent share

    One agent multiplies a person's output by s. More agents help with diminishing returns, because every agent's work still passes through one person for direction and review, and agents working side by side step on each other. C(a) describes that curve. With my starting values it rises, peaks at about four agents per person, and then falls.

  4. Account for the team

    Every added person brings handoffs and shared decisions, so γ adds a small cost per extra teammate. A bigger team finishes sooner on the calendar while the total hours go up a little.

What each symbol means

SymbolMeaningStarting value
HConventional estimate in hoursper project
fShare of the work agents can carry out95%
sLeverage from one agent: hours of conventional work delivered per hour spent directing it9
aAgents each person directs4
C(a)Effective agents per person once review and overlap are counted2.11 at a = 4
αReview bottleneck: how much agent output waits on one person0.10
βAgent overlap: conflicts and drift between agents working in parallel0.05
PPeople on the team1
γCoordination cost per extra person0.08

What the formula tells you

There is a sweet spot

Past about four agents per person, the person becomes the bottleneck and extra agents slow things down. Adding agents past that point is the wrong lever.

There is a ceiling

If 5% of a project needs human judgment, the work can never go more than 20 times faster, however many agents are running. Clear requirements raise that ceiling more than anything else.

People buy speed

Three people with agents cost about 16% more than one and finish in well under half the time. When a deadline matters, that trade is usually worth making.

The starting values reflect my own experience running two to five agents at a time, so treat them as a starting point. The share agents can handle changes with the kind of work, and the numbers get more accurate the more real projects they are checked against.

Where the numbers come from

The structure borrows from three older ideas: Amdahl's law for the split between human and agent work, Neil Gunther's Universal Scalability Law for the agent curve, and Brooks's law for team coordination. This section explains each one, where every input comes from, and what each output assumes.

The models

Amdahl's law

Gene Amdahl showed in 1967 that speeding up part of a job only helps so much. The part that stays slow sets a limit on the whole job.

In this formula, agents speed up the share f, and the human share 1 − f stays as slow as it was. So no number of agents can make a project faster than 1 ÷ (1 − f). That limit is the ceiling in the calculator.

ceiling = 1 ÷ (1 − f)

Universal Scalability Law

Neil Gunther introduced the Universal Scalability Law in 1993 to model how systems scale under load. It starts from ideal parallel work and subtracts two costs. Contention, α, is time spent waiting in line for a shared resource. Coherency, β, is time spent bringing separate work back into agreement.

I apply the same curve to one person directing a agents. The person is the shared resource, so waiting for review is contention. Agents that overlap or drift apart create coherency cost. Coherency grows with every pair of agents, so the curve peaks at √((1 − α) ÷ β) agents and then falls. With my starting values, the peak is about 4.2 agents.

C(a) = a ÷ (1 + α(a−1) + β·a(a−1))

Brooks's law

Fred Brooks wrote in 1975 that adding people to a late software project makes it later. New people need time to ramp up. Each pair of people also adds a line of communication, and a team of n people has n(n − 1) ÷ 2 of them.

I charge that cost as γ for each extra person. That straight line keeps the math simple for small teams. For a large team it understates the cost, so treat big-team results with caution.

team = 1 + γ(P − 1)

Where the inputs come from

The eleven inputs come from three places: your project, my own agent runs, and the costs in the models above.

From your project

Conventional estimate, H
The formula adjusts this number instead of replacing it. Get it the way you always would, from people who have built similar projects.
People, P, and focused hours per day
Focused hours count real working time, not time at a desk. The starting value is 6, not 8, because meetings and interruptions take part of every day.
Hourly rate and risk buffer
The rate is yours to set. The buffer covers estimation error, which is largest at the start of a project. Barry Boehm charted this in his 1981 book Software Engineering Economics. Early estimates can miss by a factor of 4 in either direction. Steve McConnell later named that range the cone of uncertainty (opens in a new tab). Raise the buffer when the scope is still moving.

From my agent runs

Share agents can do, f, and leverage, s
These come from my own work running two to five agents at a time, not from a published study. I start f at 95% and s at 9. Lower f for projects with a lot of discovery or older systems, because people carry more of that work. To tune both, track a few real projects against their conventional estimates.
Agents per person, a
This is how many agents each person directs at once. The USL curve, not a fixed multiplier, decides what each extra agent adds.

From the models

Review bottleneck, α, and agent overlap, β
These are the USL's contention and coherency costs. Gunther fits them by regression from throughput measured at several load levels. You can fit your own the same way, by measuring one person's output with 1, 2, 4, and 6 agents. My starting values of 0.10 and 0.05 reflect the same experience as f and s.
Coordination cost per extra person, γ
This is the Brooks cost for each added teammate. At 0.08, a team of three spends 16% more total hours than one person does.

What the outputs assume

Human hours
Human hours are the hours people spend on the project: directing agents, reviewing their work, and doing the work only people can do. The human share of H keeps its full cost. The agent share shrinks by s · C(a), and the team factor adds coordination. Agent usage costs, such as API fees, are not part of this number.
Price
Price is human hours times the rate, plus the risk buffer. The buffer is the contingency for the estimation error described under the inputs.
Calendar time
Calendar time divides human hours by the team's focused hours per day. That assumes the work splits evenly across people, which Brooks warned is rarely true. Work that must happen in order, such as a migration that blocks a launch, takes longer. Treat this number as the shortest likely schedule.