Add KCP manifest and TL;DR summaries: 76% fewer agent navigation calls

[totto]

What this adds

A knowledge.yaml manifest at the repo root, plus four pre-built TL;DR summary files for the highest-traffic documentation sections.

KCP (Knowledge Context Protocol) is a lightweight standard that tells AI agents what a repository contains and how to navigate it efficiently — a structured index with intents and trigger phrases for each documentation unit, rather than having agents explore by trial and error.

---

Benchmark: 76% fewer tool calls

8 queries, same Haiku model, tool counts from the Anthropic API (not agent self-reports):

Query	Baseline	KCP	Saved
Flows vs Crews — what's the difference?	14	2	12
Create my first agent and assign a task?	7	3	4
Create a custom tool?	8	3	5
Add memory to a crew?	7	3	4
Which LLM providers are supported?	17	5	12
Build a flow that triggers a crew?	15	2	13
Hierarchical crew with manager agent?	22	9	13
Add knowledge (RAG) to a crew?	33	3	30
Total	123	30	93

76% reduction. The standout is the RAG query (33 → 3, 91% reduction) — the baseline agent traversed the entire docs tree before finding knowledge.mdx, while KCP directed it straight to tools-memory-tldr.mdx via the rag crew trigger. The hierarchical crew query had the smallest gain (22 → 9) because answering it genuinely requires reading large reference files — KCP removes exploration overhead but cannot compress inherently dense content.

---

How it works

Without the manifest, an agent asked "how do I build a flow that triggers a crew?" typically:

1. Reads README.md
2. Globs docs/en/ directory
3. Reads introduction.mdx
4. Reads concepts/crews.mdx
5. Reads concepts/flows.mdx
6. ... 15 reads total

With knowledge.yaml, the agent:

1. Reads knowledge.yaml — sees flows-tldr unit with triggers ["flow triggers crew", "start listen", "event driven workflow"]
2. Reads docs/en/concepts/flows-tldr.mdx — gets @start/@listen/@router anatomy + a minimal Flow+Crew example in ~500 tokens
3. Done. 2 tool calls.

---

Files added

knowledge.yaml                                  ← 16-unit KCP v0.3 manifest
docs/en/getting-started-tldr.mdx               ← Flows vs Crews, minimal working crew, installation
docs/en/concepts/agents-tasks-tldr.mdx         ← Agent role/goal/backstory, task structure, process types
docs/en/concepts/flows-tldr.mdx                ← @start/@listen/@router anatomy, minimal Flow+Crew example
docs/en/concepts/tools-memory-tldr.mdx         ← Built-in tools, @tool decorator, memory types, RAG quickstart
BENCHMARK.md                                    ← Full results with methodology
benchmark.py                                    ← Benchmark runner (reproduces the numbers above)

Static files only — zero runtime overhead, no new dependencies, nothing changes for existing users.

---

KCP spec

Full specification: github.com/cantara/knowledge-context-protocol (v0.3).

This is the fifth repo we have benchmarked with KCP. Results across all five:

Repo	Type	Reduction
Application codebase (plugin wizard)	app code	74%
Documentation repo (infra agents guide)	pure docs	53%
smolagents (HuggingFace)	Python framework	73%
AutoGen (Microsoft)	Python framework	80%
CrewAI	Python framework	76%

Happy to adjust trigger phrases, add coverage for enterprise integrations or observability platforms, or restructure files to match your documentation conventions.

---

Note

Low Risk
Documentation-only additions plus an optional standalone benchmarking script; no changes to runtime library behavior, but the new script invokes local file/grep access when run.

Overview
Adds a repo-root knowledge.yaml Knowledge Context Protocol (KCP) manifest that indexes key CrewAI docs by intent/triggers and points to smaller summary units to reduce agent “search” reads.

Introduces several new TL;DR MDX pages (getting started, agents/tasks, flows, tools/memory/knowledge) and a BENCHMARK.md + benchmark.py runner to measure the reduction in tool calls when using the manifest.

^{Written by Cursor Bugbot for commit f532932. This will update automatically on new commits. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}