AI coding agents spend much of their effort looking things up. We measured what happens when an off-the-shelf agent gets its platform knowledge from GetPlane instead of the open web: the same knowledge, reshaped for machines and served from the GetPlane AI control plane.
Before a coding agent can write working code against an unfamiliar platform, it has to learn that platform's tools, field names and types. Today it does that by searching the web and reading pages written for people. We ran the same agent on the same task and changed only where its documentation came from.
Claude Code 2.1.280 with Claude Sonnet 5, default configuration, web search and fetch enabled. It pulls the platform's public documentation pages as it needs them, the way coding agents work today.
The platform documentation is reshaped into Knowledge Content Objects (NCO, patent pending) and served from the GetPlane AI control plane through its MCP reader.
The difference is shape, not content. Web documentation is organised as pages for people to browse. An NCO corpus is organised for a machine to look things up: one fixed vocabulary, an index of every exact name, and a catalogue of every tool.
Agent and model (Claude Code 2.1.280, Claude Sonnet 5), prompt, starter code, simulated backend, hidden acceptance tests, and call and time limits. A third lane with no documentation at all confirms the task can't be solved from memory.
The agent works until it declares the task done. Hidden tests it never sees then decide pass or fail. We record wall-clock time, tokens and list-price cost for every run.
In the standard setup, the median run took 13.2 minutes and $2.75. With GetPlane serving the NCO corpus and no web access, the median run took 6.1 minutes and $1.38. Both passed every hidden test.
Each dot is one complete run: the agent reads the documentation, writes and tests its code, and stops; hidden acceptance tests then judge the result. The slowest GetPlane run took 6.7 minutes; the fastest standard-setup run took 10.9.
Cost per accepted solution is the total spend of all runs divided by the runs that passed. Prices are the model provider's list prices.
Derived from the medians above: $100 / $2.77 and $100 / $1.33; 60 / 13.2 and 60 / 6.1 minutes. The agent runs one task at a time.
Median tokens per run. Each change to the corpus or the reader was pre-registered, then run with the same agent, task and hidden tests.
Web documentation is organised as pages for people. The GPX corpus format uses one fixed vocabulary, an exact-name index, and a catalogue of every tool. The GetPlane MCP reader returns one tool at a time. The run logs show the agents working in four steps.
A list of every tool and what it does. Every agent read it first or near the start (nco_open).
Field and type names come from an index instead of guesswork (nco_find).
Its input, output and the types it uses, in one call. No read was truncated in any of the six v4 runs (nco_get).
The agent produced 57% fewer output tokens than the standard setup: less drafting and rework.
With no documentation at all, the same agent passed once in three runs. The two failures used their full 300-call budget trying request shapes against the API.
Corpus v4 carries the whole reference in 2.25 MB against 5.05 MB. Every field reference resolves, every exact name is indexed, no string is lost, and a rebuild comes out byte-identical.
With the same v4 corpus, reading through the GetPlane MCP reader instead of opening the files directly used 13% fewer tokens and cost 8% less per accepted solution.
Every GetPlane run reached one network host, the model API. The documentation stayed on the machine, which matters for teams that cannot let agents browse the internet.
The NCO corpus format was developed on the retrain task above. To test it independently, we drew a new task at random from tools the format was not shaped around: an audit of a managed prompt library against policy documents. We ran it seven times in each setup: the standard setup with web access, the same agent with no documentation, and GetPlane’s NCO corpus through the MCP reader and as plain files, 42 runs in all.
Cards compare the standard setup with web access against NCO v4 through the GetPlane MCP reader, 7 runs each, medians. The GetPlane runs were made one day before the web and no-documentation runs, with the same agent, model, task and hidden tests. The web agent read Google’s public documentation; GetPlane runs reached only the model API. A gain the same size as on the retrain task shows the earlier result was not an artefact of developing the format on that task.
On a task that requires documentation, an AI coding agent working from GetPlane's NCO corpus and reader produced the same accepted result as the same agent in its standard setup with web access, in 54% less time, with 54% fewer tokens and at 52% lower cost per accepted solution.