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. In the transcript she says that three years ago she moved into a full-time role immersed in AI coding and AI on software teams, helping Thoughtworks colleagues and clients and writing about the space, including on Martin Fowler's website. The speech-to-text source renders her name and organization inconsistently; preserve those artifacts when quoting.

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)

Böckeler argues that the interesting work has moved from model-watching alone to understanding the whole coding-assistant system: model capabilities, harness features, context/harness engineering, guide-and-sensor feedback loops, and the risks of escalating autonomy. Her practical frame is not "use more AI everywhere" but match task complexity, context, harness capability, and supervision level through risk assessment while watching the human and organizational costs.

Section TOC

Terminology Glossary

• Model learning map -- The four things users need to understand: models are not magic, statelessness, context-window/attention trade-offs, and choosing the right model for the task.
• Coding harness / coding agent -- The layer around the model that provides prompts, tool integrations, code search, orchestration, UI, extensibility, and observability.
• Harness engineering -- Böckeler's working term for context engineering for coding agents: expanding the harness with codebase-specific context, skills, MCP servers, sub-agents, plugins, hooks, and similar features.
• Guides and sensors -- Her conceptual model for harness engineering. Guides feed information and constraints forward; sensors provide feedback so the agent can self-correct.
• Inferential sensors -- LLM-based checks, such as code review agents judging another LLM's work.
• Computational sensors -- Deterministic or CPU-style checks, such as static analysis, lint rules, import scanners, tests, and logs.
• Continuous drift detection / garbage collection -- Scheduled reviews for accumulating debt or risks that need not run on every CI build.
• Probability / impact / detectability -- The risk-assessment triad Böckeler uses to decide how much autonomy and supervision are appropriate for a coding-agent task.
• Cognitive X / cognitive surrender -- Her umbrella for cognitive load, cognitive debt, cognitive deferral, and the danger of surrendering active thinking to AI.

Named Frameworks / Concepts

1. The model-user learning map: not magic; statelessness; context window vs attention; which model for which task.
2. Model-task matching: reflect on files involved, blast radius, uncertainty, context size, reasoning level, and tool-calling needs before choosing a model/workflow.
3. Harness feature model: prompts, tool integrations, code search, orchestration/sub-agents, caching, UI, extensibility, and observability.
4. Harness engineering = context engineering for coding agents: use harness features to provide codebase- and workflow-specific context.
5. Guide/sensor model: combine feed-forward guidance with feedback sensors to improve first-pass quality and trigger self-correction.
6. Sensor placement along the path to production: coding session, pre-commit/integration, PR process, CI, scheduled drift detection, and production observability.
7. Autonomy risk assessment: probability, impact, and detectability determine whether the human stays in the loop, on the loop, or moves out of the loop.
8. Cognitive surrender checklist: watch for surrendering review, architectural understanding, junior development, cost discipline, sandboxing, and organizational learning.

Open Questions / Not Covered

• The transcript does not provide a detailed benchmark table comparing specific models.
• It names several tools and trends, but does not provide setup instructions for any specific harness.
• It does not prescribe one universal review policy; it says supervision depends on risk, task, context, and feedback loops.
• It references external posts, newsletters, papers, and talks, but this bundle only grounds answers in the transcript itself.