Your typing speed isn't the bottleneck.

The model

Each activity in your week (coding, review, debugging, meetings, context switching, glue) carries two coefficients:

• α (alpha) — fraction of time saved by AI as a typing assistant: a single human, working serially.
• β (beta) — additional fraction unlocked by coductor-style orchestration of parallel agents on top of α.

This split is the load-bearing design choice. It models two distinct curves: AI-as-autocomplete plateaus around 1.4×, but parallel orchestration keeps scaling until coordination overhead bends it down.

Coefficients per activity

The parallelism curve

Output as a function of how many parallel work streams (P) you can realistically supervise:

output(P) = P / (1 + 0.15 × (P − 1))

Same hyperbolic-saturation family as Amdahl's law. The 0.15 constant captures real-world coordination overhead per added stream. Sample values:

The curve saturates around 5×, honestly bending past P=5 because supervising more than five agents in flight is a coordination problem, not a typing problem.

Team mode

When you toggle team mode, the per-seat math runs once and is summed across the team, with an organizational coordination tax applied:

team_multiplier = N × (1 / (1 + 0.05 × (N − 1)))

Bigger teams pay a small organizational tax — alignment meetings, cross-team coordination, communication overhead. So a 10-person team's leverage isn't 10× a single person's; it's closer to 7×. Larger teams compound the tax further.

Annual savings

Putting it all together for a single seat:

autocomplete_savings = Σᵢ (hours × allocᵢ × αᵢ) × rate × 50_weeks
coductor_extra       = Σᵢ (hours × allocᵢ × βᵢ) × output(P) × rate × 50_weeks
leverage             = coductor_extra × team_multiplier
weeks_gained         = leverage / (rate × 40)

We use 50 working weeks per year (allowing for holidays and PTO) and a 40-hour week to convert dollars back to "weeks of working time." The headline number is leverage — what you gain above the AI-as-autocomplete plateau, not the autocomplete savings themselves.

What this doesn't capture

• Quality differences. AI can write code faster but may introduce subtle bugs. The α coefficient discounts for this, but the discount is necessarily approximate.
• Onboarding cost. There's a real learning curve to running parallel agents productively. Year-one returns are typically lower than the steady-state numbers shown here.
• Review and verification scaling. If agents produce more code, someone has to review it. Throughput isn't free downstream.
• Infrastructure cost. Running multiple agents has compute, tooling, and observability costs not modeled here.
• Domain variance. A research codebase, a regulated medical product, and a CRUD app behave very differently. Defaults are population averages.

Sources

• DORA — DevOps Research and Assessment, including the State of AI-Assisted Software Development reports.
• Stack Overflow Developer Survey — annual snapshot of where developers actually spend their time.
• Atlassian State of Developer Experience — quantified friction points and time-loss research from 3,500+ developers and managers.
• Gloria Mark's research on attention recovery — longitudinal studies on context-switch cost and focus recovery.
• METR 2025 — independent measurements showing real-world AI productivity gains track lower than vendor claims (in fact, experienced devs were ~19% slower with AI assistance), motivating the conservative α coefficients above.