WebinarNotes/wiki/entities/thorsten-ball.md

Thorsten Ball

#entity

Summary

Founding engineer at amp (Sourcegraph's agent product); author of Writing an Interpreter in Go and Writing a Compiler in Go. The corpus's most extreme practitioner datapoint: 99% of the code at his company is written by AI, and he uses no skills, no MCP servers, and no custom slash commands.

Current Understanding

Thorsten's position is that the frontier has moved past tooling questions. What matters is (1) where the agent's information comes from, (2) how agent-friendly your codebase and workflow are, and (3) knowing what you want to build. Everything else — model choice, effort levels, prompt tricks — he treats as diminishing returns or hobby.

His operating discipline is shedding-weight: aggressively deleting processes and products that only made sense when humans were the bottleneck. AMP publicly kills its own features (the VS Code extension: "who has the editor open anymore?"), and he applies the same knife to backlogs, CI that repeats the agent's tests, and admin panels.

He is also the corpus's clearest voice on taste surviving automation: 99% AI-written code plus 15 AI-generated icon variants he picks between is not slop, because "slop comes from humans not having good product." This is theo-browne's position stated from inside a company rather than from a podium — the two are the corpus's closest allies.

Where he cuts against the vault: his answer to "how does the agent know things?" is a good prompt plus a good codebase plus AGENTS.md — not skills-as-memory. He talks to the model "like a Slack message to a senior engineer": set the standard, state intent, riff on design, specify process, set sub-agent economics. Worth holding as a live alternative rather than dismissing: he is a professional engineer working daily in one codebase he controls, which is the situation where codebase-as-context works best and portable skills matter least.

Evidence

• Interview "Agentic Engineering, explained by a 10x developer" (42:33, with David Andre): shed weight, orbs, Emacsification, internal software, first-principles thinking, prompt structure, predictions — 2026-07-28-agentic-engineering-10x-developer.
• Built a club food-ordering app from a menu photo in 3 × 5-minute iterations; encoded a 20-person club's ordering process in ~2 hours of phone typing.
• Forked the diff viewer hunk and had AMP add a Gruvbox theme + sidebar file-checkoff in two minutes of agent time; never upstreamed.
• Shipped a change to production live during the podcast recording.

Related Pages

• Tools/orgs: amp
• Concepts: shedding-weight, build-for-the-agent-not-the-human, emacsification-of-software, explosion-of-internal-software, async-by-default, code-as-throwaway, context-as-scarce-resource, product-ownership
• Compare: theo-browne (nearest ally — same slop-is-human, same code-is-cheap stance), konstantin and allie-miller (both build the skills layer he skips), sebastian (the enterprise reality that resists "kill your process")

Contradictions / Uncertainty

• The skills rejection. "No custom slash commands, no skills, no MCP servers" is a direct counter to the vault's machine-side spine (skills-as-memory, evolution-of-agent-tooling). Unresolved; likely scoped to his situation, but he does not scope it himself. Status: tentative.
• Model choice doesn't matter. He says stop tuning between models; eugene deliberately runs 4.7 over 4.8 for over-proactivity (leave-less-room-for-imagination). Status: tentative.
• The 99%-AI-written figure and the team poll are self-reported from inside a company that sells an agent. Status: tentative.
• "Local dev is going away" is a prediction from someone selling remote sandboxes.
• The remix/internal-software predictions now have a sourced counter. 2026-07-29-what-if-we-vibe-code-it argues the cost of software is maintenance-is-the-real-cost, and offers an observed reversal (in-house Jira clone → back to Linear in four months). His club app passes that source's build-vs-buy checklist; his "teams will remix Riverside" prediction is exactly what the checklist rejects. Status: tentative on both sides — one anecdote against one prediction.

Next Questions

• Would he still skip skills if he worked across many unrelated codebases, or as a non-engineer with no codebase to encode context into?
• What is actually in AMP's AGENTS.md, and how is it maintained? That file carries the entire load his setup places on it.