Coding agent
An AI agent dedicated to assisting with or autonomously completing software development tasks such as writing, debugging, and refactoring code. It is a core form of AI programming tools and can be integrated into an IDE or run standalone.
Core Facts
Timeline (last 90 days)
A pragmatic starting recommendation: first use the coding agent's CLI mode to get a small end-to-end project running, then introduce an orchestration framework to standardize the loop and tool calls; finally, consider advanced features such as multi-agent collaboration and long-term memory.
The capability boundary of coding agents is expanding from generating code to verifying multimodal output, and audio is the next perception gap after vision.
Coding Agents, research Agents and customer service Agents are already common Agent application scenarios.
Traditional front-end design review, feedback collection, and item-by-item code modification take weeks, which shrinks to days after adopting coding agents, while the perception agent workflow can compress the cycle to hours.
In coding agent scenarios, task state, conversation history, and tool-calling protocols are often bound to a specific agent, making the migration cost of switching far higher than swapping an HTTP endpoint.
In a real coding session case, the main Agent's context grew from about 60,000 tokens to 400,000 tokens.
A coding Agent is essentially a loop system making decisions in a large action space, where at each step it must choose which file to read, which line to change, and when to stop.
Large build logs are close to the 99th percentile in the observation distribution, belonging to extreme cases rather than everyday normality.
An analysis of 600 public coding agent trajectories found that the number of turns explains 84% of the variance in log trajectory cost, while the average observation size explains only 4%.
Among the 600 trajectories in this study, 98.6% of input tokens were provided by the prefix cache.
16 more timeline events
All Facts (20)
In coding agents, the number of turns is often jointly determined by the complexity of the task itself, the size of the codebase, and failure retries, and directly cutting turns may shift the cost onto the success rate.
75%Unverified假设企业每天运行5000次Agent任务、每次平均消耗$0.50 API费用，5%效率提升意味着年化节省超过$45,000
90%UnverifiedA pragmatic starting recommendation: first use the coding agent's CLI mode to get a small end-to-end project running, then introduce an orchestration framework to standardize the loop and tool calls; finally, consider advanced features such as multi-agent collaboration and long-term memory.
50%UnverifiedCoding Agents, research Agents and customer service Agents are already common Agent application scenarios.
50%UnverifiedTraditional front-end design review, feedback collection, and item-by-item code modification take weeks, which shrinks to days after adopting coding agents, while the perception agent workflow can compress the cycle to hours.
50%UnverifiedIn coding agent scenarios, task state, conversation history, and tool-calling protocols are often bound to a specific agent, making the migration cost of switching far higher than swapping an HTTP endpoint.
50%UnverifiedIn a real coding session case, the main Agent's context grew from about 60,000 tokens to 400,000 tokens.
50%UnverifiedA coding Agent is essentially a loop system making decisions in a large action space, where at each step it must choose which file to read, which line to change, and when to stop.
50%UnverifiedLoop engineering for coding agents encompasses context management, model selection, compression strategies, and termination conditions.
50%UnverifiedThe study derives and empirically validates a closed-form break-even condition for model delegation, finding that in half of the evaluation settings delegation instead increases cost.
50%UnverifiedThe study covers two agent scaffolds and two language models, totaling 600 public coding agent trajectories.
50%UnverifiedThe study does not prove that turns can be removed while keeping the task success rate unchanged; the evidence is correlational rather than causal.
50%UnverifiedAn analysis of 600 public coding agent trajectories found that the number of turns explains 84% of the variance in log trajectory cost, while the average observation size explains only 4%.
50%UnverifiedAmong the 600 trajectories in this study, 98.6% of input tokens were provided by the prefix cache.
50%UnverifiedEstimating loop cost at flat input prices, under near-ideal caching conditions it was overestimated by about 7 times on average.
50%UnverifiedIn this study, the median observation size was only 118 to 286 tokens.
50%UnverifiedLarge build logs are close to the 99th percentile in the observation distribution, belonging to extreme cases rather than everyday normality.
50%UnverifiedIn the work of running coding agents on self-hosted open-source models, choosing the model is only the easiest part of the whole engineering effort.
50%UnverifiedOver the past year, the most mature Agent deployment scenario in the industry has been coding Agents, because programming tasks have clear verification criteria and standardized tool interfaces.
50%UnverifiedEveryday Agents for ordinary consumers are more difficult than coding Agents, as task boundaries are vague and involve cross-application operations.
50%