Every link from the talk.

• ronaldashri.comHome page, publications, talks.
• Agent-Based Software DevelopmentMichael Luck, Ronald Ashri and Mark d'Inverno. Artech House, 2004. Methods, tools and standards for building software out of agents.
• The AI-Powered WorkplaceRonald Ashri. Apress, 2019. The 2035 scenario on the hidden slides is the book's epilogue.
• OpenDialogConversational agents for insurance, with a focus on trust and safety.
• Spec KittyThe team whose experience running agents on real work informs much of this talk. See also the tools section below.
• Agents DecodedThe newsletter, on Substack.

One link per date on the abstraction ladder.

1949 · Assembly

• Wheeler's Initial Orders on EDSACThe first working assembler, Cambridge.
• EDSACFirst ran in May 1949.

1957 · High-level languages

• Fortran, Backus at IBM, 1957
• The Development of the C LanguageDennis Ritchie's own history. Bell Labs, 1972.

1995 · Objects, managed runtimes

• Java released by Sun, 23 May 1995
• Python 0.9.0, February 1991
• C++ first commercial release, 14 October 1985

2004 · Frameworks, open source

• Rails open-sourced, July 2004
• Django public, July 2005
• jQuery announced at BarCamp NYC, 14 January 2006
• npm first release, 12 January 2010

2006 to 2014 · Cloud, infrastructure as code

• EC2 beta, 24 August 2006
• Docker, March 2013
• Kubernetes announced, June 2014
• Terraform 0.1, 28 July 2014

2023 · AI pair programmer

• GitHub CopilotTechnical preview 29 June 2021, general availability 21 June 2022.
• Copilot Chat general availability, December 2023
• Cursor, Anysphere, March 2023

2025 · Coding agents

• Claude Code timelineResearch preview 24 February 2025, general availability 22 May 2025.
• Codex CLI, 16 April 2025
• Codex cloud agent research preview, 16 May 2025
• GitHub Copilot coding agent, 19 May 2025General availability 25 September 2025.
• DevinUnveiled 13 March 2024, general availability December 2024.

2026 · Fleets of agents

• No external claim. This is the talk's own framing.

Grouped by the layer on the slide, and the loop step each layer serves.

Run many at once Work

• ConductorMelty Labs. Parallel Claude Code sessions on a Mac, one worktree each.
• Vibe KanbanBloop. Open source board for running several agents side by side.
• Claude Code agent teams and subagentsAnthropic's documentation.
• Cursor cloud agentsFormerly background agents. Up to eight in parallel on one task.

Backlog in, pull requests out Work

• Gas TownSteve Yegge. Twenty to thirty agents with a coordinator. Open-sourced 1 January 2026.github.com/gastownhall/gastown
• SymphonyOpenAI's own backlog-to-pull-request runner.
• gh-aw, GitHub agentic workflowsWorkflows written in Markdown, compiled to GitHub Actions.
• claude-code-action and codex-actionThe vendors' GitHub Actions.
• Ralph loopsAn agent looping over a task list until it exits. The simplest fleet there is.github.com/umputun/ralphex

Specification ShapeWork

• GitHub Spec KitMIT licensed. Bolts on to any agent.
• KiroAWS. Turns a prompt into requirements before it writes code.
• Spec KittyOpen source, MIT. Governance, decision moments, a record both people and agents read.github.com/spec-kitty/spec-kitty

Memory and record Know

• BeadsSteve Yegge. What to do: issues as JSON lines in git, so an agent remembers its tasks between sessions.
• cartastudio.aiHow to do it, and why: ratified, sourced decisions as a graph an agent traverses.

The case: Stripe

• Minions: Stripe's one-shot, end-to-end coding agentsStripe's own account. Over 1,300 merged pull requests a week with no human-written code.Part 2
• InfoQ on Stripe's autonomous coding agentsMarch 2026 summary.

Longer lists

• awesome-agent-orchestrators
• awesome-software-factories

• AI as Normal TechnologyArvind Narayanan and Sayash Kapoor, Knight First Amendment Institute, April 2025. The essay behind the invention, innovation and diffusion lanes.
• The AI-as-Normal-Technology view of loss of control incidentsKapoor and Narayanan, 14 September 2026. The layered defences table on slide 10 is reproduced from this post.
• normaltech.aiTheir newsletter.

The shape of the loop

• The PDSA cycleThe W. Edwards Deming Institute. Know through Shape is Plan, Work is Do, Check and Learn are Study, Release is Act.

Where agents sit today

• GitHub Copilot
• Cursor
• Claude Code
• Codex

Ten thousand agents, one week

• OpenAI's announcement8 September 2026. A proof that the 3D Navier–Stokes equations can blow up in finite time, from roughly ten thousand agents over about a week.The paper, PDF
• What OpenAI's latest controversy tells us about the future of mathGrace Huckins, MIT Technology Review, 8 September 2026. Carries the Tao quote and the authorship dispute.
• Terence Tao on MathstodonThe post quoted on the slide: a premature, purely AI solution without transparency "can contaminate this process to the point where it actually becomes a net negative for the progress of mathematics as a whole."
• Controversy erupts as OpenAI claims solution to Navier Stokes maths problemABC News, 10 September 2026.
• Navier–Stokes priority controversyWikipedia's running account of the dispute between OpenAI and the Courant and Anthropic mathematicians.

How it goes wrong

• The post by @v0xium20 September 2026. The screenshot on slide 15.

• Spec KittyOpen-source delivery control plane for teams running coding agents. MIT licensed.github.com/spec-kitty/spec-kitty
• cartastudio.aiDecisions, options, and the rules that pick between them, ratified by named people and read by agents before they open a question.