WebinarNotes/wiki/sources/2026-07-28-agentic-engineering-10x-developer.md

Agentic Engineering, explained by a 10x developer — Thorsten Ball

#source

Source Metadata

• Date: YouTube interview (publication date not stated in source), 42:33 — https://www.youtube.com/watch?v=FU5_kpTAVDo
• Raw path: raw/sources/Agentic Engineering, explained by a 10x developer.md
• Source type: podcast/video interview (conclusions doc)
• Speaker: thorsten-ball — founding engineer at amp (Sourcegraph); author of Writing an Interpreter in Go / Writing a Compiler in Go
• Interviewer: David Andre
• Ingestion date: 2026-07-28

Core Claims

• The interesting variables moved. Not "which model" or "how do I read every line," but: where does the information the agent needs live, how agent-friendly is your codebase/workflow, and what do you actually want to build. "The dominant variable in output quality is now the information you put in."
• Shed weight. Kill anything that only made sense before agents — backlogs, CI that re-runs the agent's own tests, IDE extensions, admin panels, local dev. AMP kills its own features publicly and calls itself "AMP Frontier Corporation." See shedding-weight.
• 99% of AMP is written by AI, and that is compatible with high taste. "Most of slop comes from humans not having good product. With AI they can just build trash products faster." Slop = lack of ideas and playfulness, not an AI defect.
• Build for the agent, not the human. No human should fill out forms; anything a human can do on your site an agent should be able to do; ideally bring your own agent. See build-for-the-agent-not-the-human.
• Software becomes bespoke by two mechanisms: remixing existing software you don't upstream (emacsification-of-software) and building internal tools that used to be an Excel file (explosion-of-internal-software).
• Async by default. An orb — a remote sandbox tied to one conversation — packages thread + agent + computation + diff in one shareable URL. Delegate, do something else, and ask for proof because you're waiting anyway. See async-by-default.
• He uses no skills, no MCP servers, no custom slash commands. The only thing that matters to him is where the agent gets its information: training data (lossy, stale) plus the context window (your prompt, the codebase, AGENTS.md). This directly contradicts the vault's skills spine — logged below and on skills-as-memory.
• First-principles thinking is now the top skill. "Everyone becomes an architect"; the value is seeing the workflow underneath the request and knowing solutions from other industries.

Key Evidence / Details

• Model choice: once you have Fable 5 / GPT-5.6 Sol or equivalent, diminishing returns on which you pick, and on the effort level (medium vs high vs ultra). "If you're mad your model doesn't use camelCase, rethink your software engineering, not the model." AMP's default is medium.
• AMP's setup: installed as a PWA from ampcode.com; a low/medium/high/ultra dial mapping to model + sub-agent choice; sub-agents Oracle (reviewer/advisor) and Painter (images); meta-agent Puck that controls other agents, spawns orbs, messages them, runs flows. GPT, Anthropic and GLM models all supported. Agent-to-agent communication shipped "last Friday."
• The hand-coding poll: Thorsten polled the team with options 99%+, 90–99%, <90% — he didn't expect anyone under 80%. The one engineer who said "I still write a bunch by hand" landed at ~95% when pushed.
• Taste at AI speed: he had AI generate 15 versions of the orb icon (Braille characters, different styles, 18 palettes) and picked one. AMP news imagery came from turn-by-turn Midjourney with two colleagues while reading Moby Dick.
• The admin panel that dies: he built a food-ordering app for a local club from a menu photo in 3 × 5-minute iterations; the agent also built an admin UI for prices and spelling. "I'm never going to open that. I'll just send another photo and say 'fix the pricing.'" A lot of admin UI existed only so that no code had to change.
• The remix: forked a diff viewer called hunk, told AMP "add Gruvbox dark hard theme, add file-checkoff in sidebar, compile, drop it in ~/bin" — two minutes of agent time, no reason to upstream.
• Internal software: at his 20-person club he encoded the ordering process in ~2 hours of phone typing. Two variables separate winners from losers: knowing how to use agents, and having the token budget to do it.
• The printer anti-example: asked to build an app so a tablet prints a paper receipt for the kitchen, he pushed back — "Why do you need a printer? Why not a second tablet?"
• Don't benchmark against the 1%: Mitchell Hashimoto (Ghostty, GPU-accelerated terminal emulator) is cited in every online debate, but most software is CRUD, "MySQL and something-something," which agents handle fine.
• His prompt structure (porting Puck to the CLI): set the standard ("look at how it's implemented in web UI") → state intent → riff on the design → give explicit process ("research, document, sit down and think, compile what you learned, then come up with a good idea") → set sub-agent economics ("Fable is expensive, it scares me — use GPT models for the implementation"). "This is how I would talk to a senior engineer. This is the Slack message I'd send."
• Velocity: shipping velocity up in 4 weeks; designer Tim "never fixed so many paper cuts"; screenshot a bug → orb returns a fix → spot check → merge. A demo change was shipped to production live during the podcast.
• Predictions: local dev goes away; model distinctions matter less ("a button with which you can spawn a John Carmack"); unclear what software survives remixability; infra margins get eaten (15+ sandbox providers racing to zero); everyone moves up a level of abstraction.

Connections

• Entities: thorsten-ball (new) · amp (new) · claude-code (peer harness) · theo-browne (closest ally in the corpus)
• Concepts created: shedding-weight · build-for-the-agent-not-the-human · emacsification-of-software · explosion-of-internal-software · async-by-default
• Concepts reinforced: context-as-scarce-resource (information > tuning; the token-budget variable) · code-as-throwaway (99% AI-written at a real company) · make-more-cheap-code (15 icon variations; ask for proof) · product-ownership (first-principles thinking, everyone an architect) · harness (AMP as a second reference harness; Oracle/Painter/Puck) · seniority-and-the-junior-squeeze (what took 2–3 years to teach is now a 30-second output)
• Concepts contested: skills-as-memory and evolution-of-agent-tooling (he skips the whole progression) · leave-less-room-for-imagination (model choice as a lever — he says stop tuning; Eugene selects 4.7 over 4.8)
• Related sources: 2026-07-24-youre-reading-way-too-much-code and 2026-07-14-everything-we-knew-about-software-has-changed (Theo — same strategic register, same slop-is-human stance); 2026-07-14-sebastian-eugene-interview (the enterprise counterweight to "kill your process"); 2026-07-21-larysa-interview (his AGENTS.md-only approach is exactly what leaves Larysa's memory gap unsolved)
• Named but not in the vault: Mitchell Hashimoto, Ghostty, hunk, Sourcegraph, Midjourney, Cloud9, Quinn (AMP CEO), Riverside

Open Questions

• Does "no skills, no MCP" generalize, or is it a property of his situation? He works daily in one codebase he controls, with a company harness he helped build and a team-maintained AGENTS.md. The vault's skills case is strongest for people who move across many ad-hoc tasks and cannot encode context in a codebase (HR, BA). Neither side is tested against the other. Status: tentative.
• What replaces the token budget as a constraint? He names it as one of two winner/loser variables but says nothing about who pays for it — the missing economics of explosion-of-internal-software.
• All ratios (99% AI-written, the team poll) are self-reported from inside the company that sells the agent. Status: tentative.
• If admin panels and forms die, what does a non-technical person operate? Thorsten's answer is "prompt the agent," which assumes the prompting skill the corpus's HR interviews say is the actual bottleneck (levels-of-ai-usage).
• "Local dev is going away" is a prediction from a company selling remote sandboxes, and sits against eugene's consolidated local workspace pitch (harness). Status: tentative.

Change Impact on Wiki

• Created 5 concepts: shedding-weight, build-for-the-agent-not-the-human, emacsification-of-software, explosion-of-internal-software, async-by-default.
• Created 2 entities: thorsten-ball, amp.
• Updated skills-as-memory and evolution-of-agent-tooling with the corpus's first credible rejection of the skills abstraction (recorded as a contradiction, not smoothed).
• Updated context-as-scarce-resource (information > tuning; token budget), code-as-throwaway (99%-AI-written datapoint; slop-is-human), make-more-cheap-code (variations-not-answers; ask-for-proof), product-ownership (first-principles as the top skill), harness (AMP; the minimal-harness position), seniority-and-the-junior-squeeze (the collapse of hand-taught senior knowledge), enterprise-ai-reality (token budget as a new access divide), leave-less-room-for-imagination (his prompt structure as a worked example; model-tuning tension), personal-ai-operating-system (the fourth layer — tools you build for yourself), claude-code (AMP as peer harness), theo-browne (ally cross-link), overview (9→10 sources; new "frontier side" of the through-line), index.md.