coding agents
Core Facts
Timeline (last 90 days)
The per-task cost of coding agents can increase by up to three times, mainly because the system prompt is sent and accumulates with every step.
Well-known developer Lynch raised the question that coding agents perform poorly mainly not because of the model itself, but because of the external framework (harness) built for the agent.
As model capabilities gradually converge, harness quality will become the core competitiveness that distinguishes excellent coding agents from mediocre ones.
The effectiveness of a coding agent is the product of the model and the harness; ignoring the harness and purely pursuing model upgrades is like putting a stronger engine in a race car while ignoring the gearbox and tires.
For coding agents, if actions such as file writes, patch application, branch creation, and PR submission lack idempotency guarantees, retries can lead to duplicate insertions, conflicting overwrites, or redundant commits.
In coding agent scenarios, a context window that is too small will miss key dependencies and cause changes to break other modules; a window that is too large introduces irrelevant files, dilutes attention, and drives up cost and latency.
A coding agent's harness includes prompt templates, tool-calling interfaces, context management, error recovery mechanisms, and feedback loops; it determines how the model perceives the task, performs actions, and learns from results.
Simon Willison acknowledges that remarkable things can be accomplished with coding agents, but that unlocking their full potential requires extraordinary discipline and knowledge.
Simon Willison believes that the more deeply one uses coding agents, the more convinced one becomes that these tools make software engineering harder rather than easier.
Almost all mainstream AI coding assistants today come from U.S. vendors, and Europe lacks mature direct alternatives locally.
6 more timeline events
All Facts (16)
The simultaneous maturation of low-cost robots and coding agents is lowering the overall barrier to robot development.
65%UnverifiedThe per-task cost of coding agents can increase by up to three times, mainly because the system prompt is sent and accumulates with every step.
50%UnverifiedThe effectiveness of a coding agent is the product of the model and the harness; ignoring the harness and purely pursuing model upgrades is like putting a stronger engine in a race car while ignoring the gearbox and tires.
50%UnverifiedWell-known developer Lynch raised the question that coding agents perform poorly mainly not because of the model itself, but because of the external framework (harness) built for the agent.
50%UnverifiedA coding agent's harness includes prompt templates, tool-calling interfaces, context management, error recovery mechanisms, and feedback loops; it determines how the model perceives the task, performs actions, and learns from results.
50%UnverifiedIn coding agent scenarios, a context window that is too small will miss key dependencies and cause changes to break other modules; a window that is too large introduces irrelevant files, dilutes attention, and drives up cost and latency.
50%UnverifiedFor coding agents, if actions such as file writes, patch application, branch creation, and PR submission lack idempotency guarantees, retries can lead to duplicate insertions, conflicting overwrites, or redundant commits.
50%UnverifiedSimon Willison believes that the more deeply one uses coding agents, the more convinced one becomes that these tools make software engineering harder rather than easier.
50%UnverifiedSimon Willison acknowledges that remarkable things can be accomplished with coding agents, but that unlocking their full potential requires extraordinary discipline and knowledge.
50%UnverifiedAlmost all mainstream AI coding assistants today come from U.S. vendors, and Europe lacks mature direct alternatives locally.
50%UnverifiedWhen an AI agent reads context, generates text, and may be logged in the cloud, the traditional credential isolation assumptions based on environment variables, secret management services, or .env files break down.
50%UnverifiedPilot experiments had three coding agents and eight models each run eight 20-minute single-pass tasks, with label-aware F1 showing extremely high variance ranging from 0.143 to 0.892.
50%UnverifiedA pre-defined skill file is a capability encapsulation that enables a coding agent to understand how to invoke Blender.
50%UnverifiedDevelopers can use coding agents and give them full workspace context.
50%UnverifiedThe success rate of programming agents handling complex tasks has been continuously rising.
50%UnverifiedAs model capabilities gradually converge, harness quality will become the core competitiveness that distinguishes excellent coding agents from mediocre ones.
50%