CodingSoftware & developer tools
Developer Jeremy Scheff writes an 'obituary for programming' as coding agents boost his business but take away the craft
Jeremy Scheff argues coding agents have made his software business more productive than ever while removing the puzzle-solving he loved. The essay, amplified by Simon Willison, lands amid mixed evidence on agent productivity and falling employment for young developers.
From the Simon Willison article of October 19, 2026 and cited sources: Scheff describes agents that fix bugs from logs, speed up code and finish long-delayed refactors. [2] [1] [4] [9]
SOURCES[2] Jeremy Scheff, "My obituary for programming" dumbmatter.com[1] Simon Willison, "My obituary for programming" (link post) simonwillison.net[4] Simon Willison, "Deep Blue" simonwillison.net
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Scheff describes agents that fix bugs from logs, speed up code and finish long-delayed refactors. [2] [1] [4] [9]
- Original source
- Simon Willison article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 19, 2026 article published
- Covered
- Oct 20 edition
- Evidence
- 9 sources across 6 sites
- 1Simon Willison, "My obituary for programming" (link post)simonwillison.net
- 2Jeremy Scheff, "My obituary for programming"dumbmatter.com
- 3METR, early-2025 developer studymetr.org
- 4Simon Willison, "Deep Blue"simonwillison.net
- 5METR, February 2026 uplift updatemetr.org
- 6Stanford Digital Economy Lab, "Canaries in the Coal Mine?"digitaleconomy.stanford.edu
Reach: not measured yet
Scheff describes agents that fix bugs from logs, speed up code and finish long-delayed refactors. He also mourns the loss of programming as paid problem-solving. These speedups are his own anecdotes, not measurements. Willison sees a split: developers focused on finished systems gain, while those who loved the puzzle lose something real. He links this to a 'Deep Blue' sense of ennui he has tracked since early 2026. [2] [1] [4] [9]
The productivity evidence is unsettled. In METR's 2025 randomized trial, experienced developers were 19% slower with AI but believed they were about 20% faster. METR's late-2025 follow-up pointed to a speedup, but its confidence interval included zero. The study design also broke down: some participants avoided submitting tasks where AI helps most, and running several agents in parallel made timing unreliable. [3] [5] [8]
Stanford researchers report that employment of 22–25-year-olds in AI-exposed occupations is 19% below that of less-exposed peers, mostly because of reduced hiring. The authors call this descriptive rather than causal, and the effect weakens when education is controlled for. [6]
Measured evidence shows uncertain speed gains and a drop in young-developer employment that has not been shown to be caused by AI.
Read the full assessment
Implications: track quality and outcomes, not perceived speed; watch engagement among craft-motivated engineers; protect the junior talent pipeline.
Executive brief
Jeremy Scheff says AI coding agents have made his software business more productive than ever. He still wrote an obituary for programming, because the agents removed the part he enjoyed, which was solving the problem himself (Scheff). On 19 October 2026 Simon Willison linked the essay and summed up the split: developers who care about the finished system gain, and developers who loved the puzzle lose something real (Willison). The essay is commentary, not evidence. The data around it points in two directions. Measured productivity gains are uncertain. Payroll data shows a fall in jobs for young developers that is large but has not been shown to be caused by AI.
What changed and event timeline
METR finds AI slowed experts
In a randomized trial, 16 experienced open-source developers took 19% longer on 246 tasks when AI was allowed. They still believed AI had made them about 20% faster ().
"Deep Blue" gets a name
Willison and Adam Leventhal named developers' AI-driven ennui and dread on the Oxide and Friends podcast, after the chess and Go defeats by Deep Blue and AlphaGo ().
METR's estimate shifts toward a speedup
METR's late-2025 cohort pointed toward faster work. The authors called the result weak evidence and probably a lower bound, because of selection effects ().
Stanford revises its "canaries" paper
Employment of 22–25-year-olds in AI-exposed occupations is reported 19% below the level of less-exposed peers, mainly because of less hiring ().
Scheff publishes his obituary
He contrasts quick agent-made fixes, speedups and refactors with the loss of programming as paid puzzle-solving ().
18–19 Oct 2026: The essay spreads
It is discussed on Lobsters, where commenters are divided, and Willison's link post follows (;).
Capabilities and access
- Scheff describes coding agents in general terms. The reviewed text does not reliably name one model or version.
- Willison's February post gives Claude Opus 4.5/4.6 and GPT-5.2/5.3 as the agents that reliably produce working, tested code (Willison).
- METR's two studies used different tools: Cursor with Claude 3.5/3.7 Sonnet in early 2025, then Claude Code with Opus 4.5 in its redesigned study (METR, 2025; Birchtree summary).
Read the full section
- Scheff describes coding agents in general terms. The reviewed text does not reliably name one model or version.
- Willison's February post gives Claude Opus 4.5/4.6 and GPT-5.2/5.3 as the agents that reliably produce working, tested code (Willison).
- METR's two studies used different tools: Cursor with Claude 3.5/3.7 Sonnet in early 2025, then Claude Code with Opus 4.5 in its redesigned study (METR, 2025; Birchtree summary).
Technical analysis for researchers and developers
The hard part of this debate is measurement. METR's 2025 study assigned AI use at random to each task. When METR tried to repeat it in late 2025 (57 developers, 143 repositories, 800+ tasks), the design started to break down (METR update): The practical lesson: per-task timing studies can no longer cleanly isolate what agents contribute.
Read the full section
The hard part of this debate is measurement. METR's 2025 study assigned AI use at random to each task. When METR tried to repeat it in late 2025 (57 developers, 143 repositories, 800+ tasks), the design started to break down (METR update):
- Developers who would not work without AI declined to take part.
- 30–50% of participants avoided submitting tasks where AI would help most.
- Running several agents at once made time measurements unreliable.
- Output quality differed between the AI and no-AI conditions.
The practical lesson: per-task timing studies can no longer cleanly isolate what agents contribute. Teams measuring their own gains should track quality as well as speed, and watch for bias in which tasks get counted.
Claims and evidence
- Agents fix bugs from logs, speed up code 2–3x and finish long-delayed refactors. These are Scheff's own anecdotes and were not independently measured (Scheff).
- Experienced developers were slowed down in early 2025. This comes from a randomized controlled trial (METR).
- The late-2025 speedup. For returning developers the point estimate was −18%, with a confidence interval of −38% to +9%, which includes zero (METR update).
Read the full section
- Agents fix bugs from logs, speed up code 2–3x and finish long-delayed refactors. These are Scheff's own anecdotes and were not independently measured (Scheff).
- Experienced developers were slowed down in early 2025. This comes from a randomized controlled trial (METR).
- The late-2025 speedup. For returning developers the point estimate was −18%, with a confidence interval of −38% to +9%, which includes zero (METR update).
- Young-worker job decline. The authors call the finding descriptive, not causal. It weakens when education is controlled for (Stanford).
Context and prior work
Willison has tagged nine posts "deep-blue" since January 2026, including a Thoughtworks report, "Coding After Coders" and a September post about feeling sad about AI (tag archive). Scheff uses Matrix-style images rather than the chess comparison. He calls agent-driven work "vibecoding", something like a simulation of programming (Scheff).
Read the full section
Willison has tagged nine posts "deep-blue" since January 2026, including a Thoughtworks report, "Coding After Coders" and a September post about feeling sad about AI (tag archive). The comparison to chess is the starting point: people kept playing chess after machines beat them, but what the game meant to them changed. Scheff uses Matrix-style images rather than the chess comparison. He calls agent-driven work "vibecoding", something like a simulation of programming (Scheff).
Limitations, safety and contested findings
- Scheff's essay reflects one developer's experience and does not measure anything.
- Lobsters commenters disagree:
- One warns that unchecked AI "slop" will degrade codebases.
Read the full section
- Scheff's essay reflects one developer's experience and does not measure anything.
- Lobsters commenters disagree:
- One warns that unchecked AI "slop" will degrade codebases.
- Others say junior hiring has largely stopped.
- Willison argues that experienced programmers remain necessary and points to companies hiring more interns (Lobsters).
- The Stanford paper notes that some of the trends began before generative AI and are weaker in national survey data (Stanford).
- METR describes its own recent estimate as weak evidence (METR update).
Business and practitioner implications
- Morale is a real cost. Teams whose engineers are motivated by the craft may lose engagement even as output goes up.
- Hiring cuts at entry level now could leave a shortage of experienced engineers later (Stanford).
- Don't rely on felt speed. METR's developers wrongly believed AI had sped them up. Measure results instead, including code quality and maintainability (METR).
Read the full section
- Morale is a real cost. Teams whose engineers are motivated by the craft may lose engagement even as output goes up. Scheff's case shows a company having a strong year while its founder feels the work has lost its meaning (Scheff).
- Plan for juniors. Hiring cuts at entry level now could leave a shortage of experienced engineers later (Stanford).
- Don't rely on felt speed. METR's developers wrongly believed AI had sped them up. Measure results instead, including code quality and maintainability (METR).