David Ondrej and Thorsten Ball argue that capable coding agentsAn AI coding agent is a tool-using AI system that can inspect, modify, and validate software within a repository. change the developer's main constraint. Producing code becomes easier, so the harder work is deciding what should be built, describing the desired behavior and judging whether a large generated change is actually correct and useful.
Thorsten Ball recommends giving an agent high-quality task contextContext engineering designs how relevant instructions, evidence, state and tool information are assembled for a model's task. and then allowing it to work asynchronously in an isolated environment. The developer can review several completed attemptsCode review examines proposed source changes to assess their correctness, clarity and suitability before or around integration. instead of watching each token or repeatedly steering the model. This makes proof, tests and clear success conditions more important than conversational supervision.
David Ondrej and Thorsten Ball describe a workflow in which each task receives its own temporary environment, dependencies and verification steps. Agents can investigate unfamiliar code, run testsDeterministic software verification checks an output with a repeatable procedure that produces the same result for the same inputs and state. and delegate bounded researchAI agent delegation assigns a bounded task, context, authority, and expected result from one participant to an agent or another agent. to other agents. Humans remain responsible for architecture, product tradeoffs and the final decision to accept a change.
Thorsten Ball expects software to become more personal and remixable as implementation costs fall. Existing tools can be adapted to a team's exact workflow, but fast generation does not remove the need for taste or accountability. The durable advantage is a development process that supplies good context and demands evidence before generated work is trusted.
Watch on YouTube




