Eric Zakariasson — Building your own software factory

Source index for the AI Engineer 2026 talk. Raw transcript at eric-zakariasson-software-factory-2026 (raw/transcripts/).

Thesis

The shift from "writing code" to "managing agent fleets" requires treating your codebase + infrastructure as a factory — with assembly lines, guardrails, verification stations, and supervisors. Most teams aren't there yet; Cursor is closer than most because they dogfood aggressively.

Structure

1. Levels of autonomy (Shapiro) — levels-of-autonomy-shapiro
2. Precursor to factory — prerequisites
3. Building the factory — software-factory four components
4. Running the factory — manager mindset, sync-plan-async-execute
5. Scaling the factory — isolated-agent-vms, nested management
6. Q&A — labor/comp, handoff between PMs and engineers, tracking markdown sprawl, brownouts

Concepts introduced

• software-factory — the full framing
• levels-of-autonomy-shapiro — Dan Shapiro's 6-level ladder
• emergent-cursor-rules — rules grow from failure, don't pre-install
• verifiable-systems-for-agents — how agents check their own work
• isolated-agent-vms — VM-per-agent scaling
• sync-plan-async-execute — the operating rhythm

Entities

• eric-zakariasson — speaker
• cursor — his employer + the toolchain

Notable moments

• Cursor 3 rewrite: no VS Code base anymore. Streamlined for agent-first workflow. Multi-agent orchestration is the first tab.
• Music-agent demo: Zakariasson's self-imposed constraint project — never writes code, forces discovery of what scaffolding is actually needed. Ableton-style UI, self-authored Playwright tests.
• Computer-use verification: agent records video of itself clicking through UI to verify its own work. "An AGI moment."
• Token-maxing engineer: "The 10x engineer is no longer about words per minute. It's tokens to token usage." Q&A riff on labor comp shifting from lines-of-code to agents-spawned/tokens-used.
• Brownouts/blackouts Q: humans remain accountable. Critical paths need human-written or dual-human-reviewed code. "It's easy to vibe-code close to the sun and fly too close."

Cross-ingest links

• Bridges andrej-karpathy's vibe-coding / agentic-engineering distinction into a concrete operational playbook — Zakariasson is doing Karpathy's "agentic engineering" in production.
• Parallels mitchell-hashimoto's view (agent-as-junior-engineer): both treat agents as delegate targets requiring clear specs + verification; both run them in parallel.
• Extends parallel-agent-competitions from Hashimoto's "race clones" to Zakariasson's "thousands of VMs" scale.
• jagged-intelligence explains the UX-vs-backend asymmetry he observes in verification.

Open questions

• How does "rules should emerge" scale past one team? Who curates the emergent-cursor-rules as a company grows?
• At what team size does nested-agent-management (managers of agents of managers of agents) break down?
• What happens to new-grad hiring and career ladders when L5–L6 is baseline? (Q&A raised but unanswered.)