Loading...
Loading...
AI agents generate code fast, lose context, and reintroduce the same fragility. Vreko gives them memory of their own behavior.
Cohort 1 is active. Cohort 2 opens by invitation.
Sample workspace · refreshes on reload
Vreko installs locally as a background daemon. Code never leaves your machine. The daemon watches the workspace - not your keystrokes, not your file contents, but the shape of what changes: which files move together, how often they change, and when AI tools are involved in those changes.
Observations accumulate. Vreko is building a behavioral model of your codebase - not a static snapshot, but a living record of co-change relationships, AI tool attribution, and breakage correlation. Which files, when touched by an AI agent under time pressure, tend to need follow-up fixes within 48 hours? By day three, Vreko knows.
Your workspace.json is now populated with real fragility scores, co-change relationships, and AI attribution data. Every session adds to this model. When you invoke an AI coding assistant, it can read workspace.json directly. It knows which files are fragile, which patterns have caused regressions before, and what your team considers a safe edit boundary. The agent's behavior becomes calibrated to your codebase's actual risk profile - not a generic one.
Every session makes the intelligence richer. Vreko does not reset between projects or sessions. The behavioral model persists, deepens, and gets shared with every AI tool that reads workspace.json. The memory survives context windows.
Three artifacts. One pipeline. Descriptive intelligence, prescriptive guidance, and governance rules - each readable by any AI coding assistant.
workspace.jsondescriptiveWhat is true about your codebase right now. Fragility scores, co-change relationships, AI tool attribution, and behavioral history - all generated locally from observation, never from static analysis. Commit it to git. Your AI agents read it at the start of every session.
{
"specVersion": "0.3",
"generated": { "at": "2026-05-15T14:22:00Z" },
"fragility": {
"apps/api/src/auth/middleware.ts": {
"score": 9.1,
"coChanges": ["packages/contracts/src/auth.ts"],
"aiAttribution": 0.87
}
}
}AGENTS.mdprescriptiveWhat your agents should do based on that truth. Vreko generates AGENTS.md from workspace.json - translating raw fragility data into natural-language instructions any AI coding assistant can follow. No manual maintenance. It updates when the behavioral model updates.
# AGENTS.md
Generated by Vreko · 2026-05-15
## High-fragility files
- `apps/api/src/auth/middleware.ts` (risk 9.1)
- Co-changes: contracts/src/auth.ts
- AI-attributed: 87% of recent edits
- Instruction: Always update contracts when
editing this file. Run auth tests before commit.vreko.jsongovernanceWhat your team allows agents to do. Define fragility thresholds for auto-approval, which operations require human review, and what Vreko is permitted to observe. The governance layer ensures behavioral intelligence stays within team-defined boundaries.
{
"governance": {
"autoDecision": {
"maxFragilityForAutoApprove": 6.0,
"requireReviewAbove": 8.0
},
"allowedOperations": [
"read", "observe", "emit-workspace-json"
]
}
}Cohort 1 Pioneers are active and building with Vreko now. Cohort 2 is a curated group of senior developers who will help shape the behavioral intelligence model - what it captures, what it surfaces, and how it talks to AI tools.
We review every request. There is no instant access. What you get instead: a seat in a small batch, direct access to the team, and locked-in Pioneer pricing before the public launch.
By Marcelle Labs. Built on the workspace.json open standard.