Amdahl's law for AI agents | Electric

Skip to content

Go to Cloud<br>Cloud<br>Go to Cloud

Return to top

← Blog

Amdahl's law for AI agents

By Kyle Mathews

Are you an LLM? You can read better optimized documentation at /blog/2026/02/19/amdahls-law-for-ai-agents.md for this page in Markdown format<br>Multi-agent systems are delivering real results. Geoffrey Huntley's Ralph Loop runs autonomous coding agents in a while-loop until every PRD item is complete. Steve Yegge's Gas Town orchestrates 20–30 concurrent agents across seven specialized roles. Cursor ships an 8-agent parallel system. The throughput gains are real — and the teams getting the most out of them share a pattern.<br>But the gains aren't automatic. Flask creator Armin Ronacher told The Pragmatic Engineer: "I sometimes kick off parallel agents, but not as much as I used to do. The thing is: it's only so much my mind can review!" The teams getting the most value share a pattern — and a law from 1967 explains exactly where the leverage is.<br>The Original Insight ​<br>In 1967, Gene Amdahl made a simple observation about parallel computing: the speedup of a program is limited by the fraction that must run sequentially.<br>His formula: speedup = 1 / (S + (1-S)/N)<br>Where S is the serial fraction and N is the number of parallel processors. The key insight isn't in the formula — it's in the limit. As N approaches infinity, speedup converges to 1/S. If 10% of your work is inherently serial, you will never exceed 10x speedup. Not with 100 cores. Not with a million.<br>This wasn't an argument against parallelism — it was an engineering guide for where to focus. The same logic applies to agents.<br>The Agent Version ​<br>The equivalent law for AI agents:<br>The maximum speedup from AI agents is bounded by 1/H, where H is the fraction of the workflow that requires human judgment.

H includes every moment the system blocks on a human: clarifying intent ("Did you mean the admin dashboard or the user-facing one?"), making judgment calls ("Should we refund this customer or offer credit?"), review cycles ("This isn't quite what I wanted, try again"), approval gates, resolving ambiguity ("The policy says 'handle escalations appropriately' — what does that mean?"), and applying taste ("This is technically correct but feels wrong").<br>If H = 40% of your total workflow time, no improvement in agent capability can ever get you past 2.5x speedup. At 50%, the ceiling is 2x. Even an optimistic 20% only gets you to 5x.<br>Human fraction (H)Max possible speedup50%2x30%3.3x20%5x10%10x5%20xH dominates the speedup equation, not agent capability. But unlike Amdahl's original serial fraction S — a fixed property of the algorithm — H isn't static. Better models do shrink some components of H. An agent that needs less clarification and makes fewer errors directly reduces human time.<br>The catch: the components that models reduce aren't the ones that dominate at scale. Model improvements shrink clarification and verification. Taste and novel decisions — the components that become dominant once you've automated the mechanical parts — are largely irreducible by better models. These ceilings are real at any given moment.<br>The question is how fast H is declining — and what you do with the gains.<br>Self-Liquidating H ​<br>The highest-leverage investment isn't making agents faster or smarter. It's changing the character of H. And the tools for doing that aren't new — tests, specs, automation, documented decisions. These are engineering best practices that have been around for decades. What's new is the return on investment. When agents can act autonomously on every encoded decision, the payoff for writing that test or updating that spec is dramatically higher than it used to be.<br>The goal isn't to minimize human involvement. It's to make it self-liquidating : every human intervention should produce the artifact — the test, the spec update, the documented decision — that makes the same type of intervention unnecessary next time. (The term comes from finance: a self-liquidating loan generates the revenue to pay itself off. A self-liquidating intervention generates the artifact that eliminates its own recurrence.)<br>A team where 40% of workflow time is human but it's all taste and strategy is in a fundamentally different position than a team where 40% is human but it's all "did you mean X or Y?" and "let me re-check this output." Self-liquidating practices convert the second type into the first.<br>This requires what I've been calling configurancy — the smallest set of explicit behavioral commitments (and rationales) that allow a bounded agent to safely modify the system without rediscovering invariants. Specs, conformance suites (automated test suites that verify behavior against a spec), documented rationale. Every implicit assumption in your system is a future human-blocking event.<br>A conformance suite that makes human review unnecessary is crystallized cognition — human judgment about correctness, encoded at the moment it was made...