Outline — State of Play: AI Coding Assistants

Speaker

Birgitta Böckeler — Global Lead for AI-assisted Software Delivery at Thoughtworks; software developer, architect and technical leader with ~20 years of experience. Three years ago she took a full-time role to be immersed in AI coding / AI on software teams, supporting Thoughtworks colleagues and clients. She writes about the space, including on Martin Fowler's website. (Transcription artifacts in the source render her name variously as "Bagita Bokela", "Bigita", "Bita", "bigita bller".)

Abstract (as provided)

The hype and momentum around AI coding assistants show no signs of slowing down. Every other week, we're urged to try a new model, a new workflow, or a new way of writing specs. This presentation takes a step back and looks at the past 12 months from a higher altitude: what are the broad shifts that have taken place, and where do we stand today? If you're deeply immersed in the space, this will help you see the forest for the trees. If you've been overwhelmed by the steady stream of weekly news and updates, this offers the cliff notes.

Thesis (synthesis)

The models are no longer the most interesting layer — the ecosystem, integrations, and second-order consequences around them are. To use AI coding assistants well, practitioners need a small but firm mental model: these systems are impressive math (not magic), they are stateless (the whole history is re-sent each turn), bigger context windows trade off against attention, and choosing the right model for the task is a skill learned by use, not by formal training.

Section TOC

Terminology glossary (speaker's own framings)

• "Not magic" — "very very impressive and very very useful math" (lines 92–95). Her reminder against treating LLMs as more than that.
• Statelessness — "the model doesn't have a session"; "every single time... our agent our harness basically sends the whole history of the conversation" (lines 110–116). Caching and tool-side optimizations exist but the underlying model is stateless.
• Context window vs. attention trade-off — "even though technically the context windows have gotten a lot bigger, um it comes with a trade-off on like how well the models are able to keep attention on all of the many instructions and all of the context that we're trying to feed them" (lines 131–136).
• Model selection — "Which model do we use for which task" (line 147). Framed as the biggest area and the one you learn by using, not in formal training.

Named frameworks / concepts

• The learning map for model users — a four-item list Birgitta walks through: (1) not magic, (2) statelessness, (3) context window / attention trade-off, (4) which model for which task. The first three are teachable in formal training; the fourth is learned by use.
• "The Opus 4.5 moment" (line 76, as transcribed) — the most notable model event of the past 12 months in her view; brought back people who had stopped trying AI coding for a while.
• Seeing the forest for the trees / "multiple forests" — her framing for the talk's purpose (lines 43–46).

Open questions / not covered (in the provided transcript excerpt)

The transcript provided cuts off at line 155 mid-sentence about model selection examples. Within what's provided, the talk does not cover:

• Specific benchmarks or evals of named models
• Detailed comparisons of agents/harnesses (Claude Code, Cursor, Copilot, etc.)
• SpecKit / Tessl / Kira details (only mentioned by the host in intro)
• The viral Martin Fowler post's actual content
• Concrete org-level adoption metrics from Thoughtworks clients
• Pricing, security, or governance
• Specs / spec-driven development as a workflow

If asked about any of the above, say the provided transcript doesn't reach that material.