Geoffrey Litt: 'Understand to participate' advice from AIE
Geoffrey Litt told developers they must understand code deeply to stay able to work with coding agents and avoid accumulating cognitive.
TL;DR
- 01Geoffrey Litt told developers they must understand code deeply to stay able to work with coding agents and avoid accumulating cognitive.
- 02Geoffrey Litt argued "Understand to participate" during his AIE talk, urging developers to maintain deep mental models of code so they can continue to work alongside coding agents.
- 03Simon Willison noted this observation in a weblog post published on 2nd July 2026 and recommended catching Geoffrey's recorded talk on YouTube.
Geoffrey Litt argued "Understand to participate" during his AIE talk, urging developers to maintain deep mental models of code so they can continue to work alongside coding agents. Simon Willison noted this observation in a weblog post published on 2nd July 2026 and recommended catching Geoffrey's recorded talk on YouTube.
What did Geoffrey Litt mean by "Understand to participate"?
He meant developers must keep a sufficiently deep grasp of the code to remain active collaborators when coding agents make large or complex changes. Geoffrey framed the problem as one of avoiding cognitive debt: as agents produce increasingly large and sophisticated changes, a developer's understanding can drift away from how the code actually works, which reduces their ability to contribute.
The practical core of his advice is straightforward: learn what the agent is doing so you can be an "active participant in the creative process." That includes building and preserving a "rich set of concepts" in your mind, which Geoffrey said is necessary "to think creatively and fluently about how to move something forward." If a developer lacks that fluency, their ability to participate is "meaningfully limited." The talk situates this guidance around collaboration with coding agents rather than replacing developer judgment.
How should developers respond when using coding agents?
Developers should treat interaction with agents as a collaborative creative process, explicitly examining the agent's actions and the concepts it applies so they can guide and extend the work. Geoffrey emphasised learning what the agent is doing to remain an active participant, which implies adopting practices that expose agent reasoning and preserve developer mental models.
The weblog post stresses two linked points: coding agents are capable of producing large, sophisticated changes, and that capability raises the risk of accumulating cognitive debt if developers stop understanding the system. Staying engaged requires deliberate effort to inspect, interpret, and internalise what the agent changed and why, rather than outsourcing comprehension to the tool.
Where can you watch the full talk and get more context?
The AIE talks were recorded, with "all 300+ of them" expected to be released over the following three weeks, and Geoffrey's talk is among those Simon Willison recommended watching on YouTube. Geoffrey also published a thread version of his talk on Twitter, offering another way to access his points.
Simon Willison's note with this summary is dated 2nd July 2026, and he flagged Geoffrey's talk as one he recommends catching as the AIE recordings become available.
Why it matters
Coding agents are moving beyond small edits to constructing larger, more sophisticated changes. That capability shifts the work away from mere instruction-following toward a partnership where developer understanding determines whether they can meaningfully steer, evaluate, and extend the agent's output. If teams allow understanding to erode, they risk creating cognitive debt that reduces future agility and control.
Geoffrey's framing reframes a common automation promise: the value of agents depends not just on what they produce, but on whether humans retain the concepts needed to participate in subsequent development.
What to watch
Watch Geoffrey Litt's AIE talk once the conference recordings are released, and read his Twitter thread for a condensed version of the points. Track the AIE rollout: the organisers recorded "all 300+" talks and those recordings were due to be released gradually over the next three weeks from 2nd July 2026.
- 2 July 2026Simon Willison posts summary
Simon Willison published a weblog note referencing Geoffrey Litt's AIE talk and the phrase 'Understand to participate.'
- From 2 July 2026 over the next three weeksAIE recordings released
Organisers recorded 'all 300+ of' the AIE talks; the recordings were expected to trickle out over the next three weeks, with Geoffrey Litt's talk available on YouTube.
Written by The Brieftide · Source: Simon Willison
The Brieftide Daily · 06:00
Briefs like this one, in your inbox every morning.
Continue reading
Browse the feed
DAILY BRIEFLLMs identify CIUs in aphasic discourse: few-shot F1 0.776-0.817
Few-shot prompting let Llama-3.1-8B, Qwen2.5-7B and Mistral-7B reach mean F1s from 0.776 to 0.817 on 16 Cat Rescue transcripts.
DAILY BRIEFLLM Architecture Workflow: Sebastian Raschka on Open-Weight Models
A practical, step-by-step method for inspecting open-weight LLM releases, from checkpoints and tokenizers to benchmarks and safety checks.
DAILY BRIEFThinkDeception: Progressive RL framework for multimodal deception
ThinkDeception on arXiv uses MLLMs, a step-by-step multimodal Chain of Thought dataset and a four-tier progressive RL trainer for.
DAILY BRIEFUncertainty Decomposition improves clarification F1 in LLMs
A prompt-based split of action confidence and request uncertainty enables clarification seeking and boosts F1 across five LLM backbones.