Web Rebuilt for 50x-Faster Agents, Humans Promote to Orchestrators
Human-speed web tools now bottleneck AI agents operating 10-50x faster and consuming 10-20k tokens/sec. Rebuild via optimized tools, agent-native primitives, and new scaffolding; humans shift to tool generalists, pipeline builders, business closers, and overseers.
Eliminate Human Affordances to Unblock Agent Speed
Current web infrastructure—paginated APIs (e.g., 100 rows), logins, dashboards, timeouts, rate limits, startup sequences—assumes human eyes and hands, processing at brain speed. This brilliant human-centric design now fails as agents reason 10-50x faster, write 20-40% of FAANG code (Anthropic: 80% via Claude), analyze earnings quicker than analysts, and hit junior-dev level 24/7 (per Jeff Dean, possibly sooner). Inference dominates 90% of datacenter power, targeting 10-20,000 tokens/sec per user—impossible for humans. Tool calls consume most wall-clock time, not inference; even infinite model speed yields only 2-3x productivity gain due to compilers, file systems, APIs, CRMs, ERPs built for humans. Agents waste cycles on pagination via MCP wrappers over human APIs.
Layered Rebuild: From Optimization to Agent-Native Primitives
Layer 1: Accelerate existing tools. JS ecosystem shifts to Rust/Go/Zig for 10x+ speed (e.g., TypeScript 7 in Go); Rust's strict compiler verifies agent-generated code (Lee Robinson: 38k-line image compressor, zero deps). Enterprise lags: Salesforce paginates, SAP batches at human speed.
Layer 2: Replace tools with agent primitives. No restarts—OpenAI's persistent containers/shells install deps once, run days via server-side compaction. Branch FS enables sub-millisecond copy-on-write branches for rapid agent iteration (try/kill experiments). Shared KV caches cut multi-agent latency 3-4x vs. text.
Layer 3: Generalize per bitter lesson—computation beats human scaffolding. New models (5x faster) make prior optimizations drag (e.g., agent framework from 30% to 60% overhead). Invest in primitives so fast (<<1s delays) they scale with any agent speed, measured in CPU ticks, not human lunch breaks. Every model gen erodes human interfaces like inspection UIs.
Humans' Four High-Value Roles in Agent Economy
Agents drive efficiency attractor; web splits: superfast agentic layer, human-speed interfaces. Humans promote to non-execution roles:
- Tool-using generalists (vibe coders+): Activate/start/complete via AI tools, direct long-running agents.
- Pipeline builders: Infrastructure pros for agent primitives, data flows, security, measurement—evolving from engineers.
- Business closers (salespeople): Build relationships, close deals—humans preferred; agents may hire them to boost close rates.
- Overseers (CEO-types): Brake on speed, team assembly, leadership, accept inefficiencies for values.
Optional 5th: Creatives (Steve Jobs-like): Envision polished experiences. Start teams with these; roles already emerging, dominant in 12-24 months.