The context interface itself is selected by agent economics
Source → Observed → Interpretation → Model implication
Giving AI agents design system context from the terminal: what we learned building a CLI
View source →Atlassian maintains one structured design-system dataset and shared search handlers, then exposes them through an agent skill, an MCP server, and a CLI.
Atlassian evaluated real front-end tasks using the same agent and model on the same questions. The changed variable was whether the skill delivered context through MCP, the CLI, CLI wrapper scripts, or bundled reference files.
Each run recorded pass rate, task time, agent-reported tokens, and tool calls; Atlassian also retained transcripts to inspect whether agents followed the intended context path.
The initial npx-based CLI added more than 15 seconds of startup overhead per query in Atlassian's monorepo. Moving it into a pre-installed internal CLI reduced that overhead to milliseconds, which Atlassian characterizes as a 95% reduction.
Trace inspection showed agents making repeated round trips. Atlassian added a batch command and improved fuzzy search, reporting a further 50% reduction in the remaining lookup overhead.
Against the MCP baseline, Atlassian reports that the iterated CLI configuration maintained quality while mean task time fell from 352 seconds to 325 seconds and mean tokens per task fell from 218,000 to 201,000, both reported as 8% improvements.
Atlassian reports CLI usage surpassed MCP usage by 60%. Its agent skill now surfaces the CLI by default and falls back to MCP when a sandbox cannot access the terminal.
The experiment makes the context-delivery interface a selectable part of the agent environment. Although each configuration exposes the same underlying design-system knowledge, invocation latency, batching, output shape, and agent instructions alter the end-to-end time and token cost of completing a task. Atlassian's decision to make the measured CLI configuration the default is direct evidence of Selection acting on infrastructure around the model, while retaining MCP as a fallback shows that accessibility constraints vary across execution environments.
SUPPORTS. Supports Selection by showing that controlled variation in an agent's context interface produced measurable differences in execution time and token use while maintaining reported quality. The resulting operational choice selected one configuration as the default and preserved another for environments without terminal access. The source therefore supports the claim that environments select agent configurations partly through their reliability and economics, not only through model capability.
What this does not establish
- The source does not disclose the benchmark size, number of repeated runs, score distributions, uncertainty, or statistical significance.
- The 8% results compare Atlassian's iterated CLI configuration with its MCP baseline; they do not establish that CLIs are generally faster or cheaper than MCP servers in other systems.
- Several changes were introduced during iteration, including pre-installed invocation, batching, compact output, search improvements, and revised skill instructions, so the source does not isolate each change's contribution.
- The reported 60% usage difference measures agent interface use rather than independently establishing higher task quality or business value.
- Atlassian reports maintained quality rather than a quality improvement, and does not publish the underlying pass rates in the article.
- The results are first-party and specific to Atlassian's design-system context, monorepo, internal CLI, benchmark tasks, and agent configuration.
Across organizations and task types, which properties of a context interface most strongly determine end-to-end agent quality, latency, and token cost once the underlying knowledge is held constant?