The Intent Layer
A context engineering system that gives every AI agent a senior engineer's understanding of your codebase—extracted by an automated cartography system and delivered in the format agents already consume.
The algorithm
Our cartography system processes your codebase and compresses institutional knowledge into dense, token-efficient bundles we call Intent Nodes.
Fractal compression
Leaf nodes summarize code. Parent nodes summarize their children, not the raw code. Each layer compresses the one below it.
Hierarchical summarization
Broad context from the root, specific detail where the agent is working. It knows the architecture before it reads a line of code.
LCA deduplication
Shared knowledge lives once at the shallowest node that covers all relevant paths. No duplication, no drift.
Progressive disclosure
Minimum context upfront. Agents drill into detail only where the task requires it. Lean budget, high signal.
How it's built
Automated cartography maps your system
Walks your codebase, chunks at semantic boundaries, runs fractal compression, and produces Intent Nodes across the hierarchy. Scales across 20 repos as fast as one.
Expert interviews capture what code can't express
The sharpest context comes from your engineers—invariants, edge cases, production lessons. We combine automated extraction with expert knowledge.
It stays fresh automatically
VCS hooks detect changes, update affected Intent Nodes leaf-first up the hierarchy. No manual maintenance. Context that compounds instead of decays.
How it's delivered
MCP server? CLI tool? Graph database? We chose the simplest possible delivery format on purpose: AGENTS.md and CLAUDE.md files—every standard harness already auto-loads them.
Every harness reads them
Cursor, Claude Code, Copilot, Codex—all auto-load them natively. Zero integration.
Versioned with your code
Live in your repo. Reviewable, diffable, tracked in git. Yours forever.
Nothing to break
No database, no API, no plugin. Files are the most durable format in software.
Sophistication is in the content
The format is deliberately simple. The value is the algorithm that generates what’s inside.
Explore a sample Intent Layer — click folders and 📘 nodes to see how it works.
How we measure
We don't ask you to trust a pitch deck. Every engagement starts with measurement on your actual code, your actual tasks.
Before
We select 10–20 representative tasks from your backlog—bug fixes, features, refactors—and run them through AI agents with no Intent Layer. We measure tokens used, time to completion, success rate, and output quality.
After
Same tasks, same agents, now with the Intent Layer loaded. We measure the same metrics. The delta is the value—token savings, speed improvement, quality lift. Tasks are judged by an LLM evaluator calibrated to your team's expectations.
No improvement, no charge. You review the methodology and agree on what “improvement” means before we start. If the eval doesn't show measurable improvement, you don't pay. You keep the Intent Layer either way.
Where it works best
The Intent Layer delivers reliable advantage on codebases above 1M tokens (~50–100K lines of code). Below that, agents can usually manage on their own.
If your system is many small services rather than one large monolith, that's actually one of the most effective situations for us. The Intent Layer connects what agents can't see across service boundaries—the shared contracts, implicit dependencies, and cross-cutting patterns that live between repos, not inside them.
Not sure how many tokens your codebase is?
Proof it works
Quantitative Trading • Chicago“What we do is a little bit weird and we see it when we use AI tools. They just get it constantly wrong.”
“I was skeptical. But it speeds up all the manual work... there's no denying that. I feel comfortable now.”
41 Intent Nodes across 3 repos • 14 data sources • 200+ features
Delivered a 200+ feature data pipeline migration estimated at 140 hours in just 10—going from babysitting one agent at a time to running 12 in parallel.
Start a Proof Pilot.
48 hours from repo access to benchmarked results. Minimum 1M tokens of source code.
$20K per 1M tokens mapped + onboarding and eval • No improvement, no charge