# Helen’s Foundry

> The strategic intent layer for AI agents.

Your agents ship fast. Do they ship the right thing?

Built for Seed–Series B software companies with 5–100 engineers shipping 10+ agent PRs a day. Helen’s Foundry gives every agent the live strategic picture - and checks each PR against that intent before it merges.

## What Helen’s Foundry is

Helen’s Foundry organizes a company's strategic context as a linked graph. Mission, vision, strategy, goals, personas, problems, architecture, projects, requirements, stories, and tickets form a traceable chain. Agents receive the current chain behind their work instead of disconnected search results.

The product is built for Seed–Series B software companies with 5–100 engineers shipping 10 or more agent pull requests per day. It is designed for CTOs and VPs of Engineering who need agent throughput to remain aligned with ratified product and company decisions.

## The problem in practice

A team ratifies an enterprise-first strategy. An existing ticket still targets self-service growth. An agent follows the ticket, opens a clean pull request, and passes every test. The implementation is correct, but its premise is no longer current.

Helen’s Foundry traces the pull request back to the ratified strategy and raises a finding before merge. The team can inspect the incompatible facts and decide which work should continue.

## What changes in your workflow

1. **Draft the decision chain:** An agent uses existing documents, issues, repository context, and work history to draft the initial graph.
2. **Ratify material premises:** An authorized person ratifies the current strategies, goals, and constraints.
3. **Give agents linked work:** Each agent receives the ticket, its current decision chain, open findings, and acceptance criteria.
4. **Check the result:** Helen grades the linked pull request against that chain. Findings go to Reconcile, where a human chooses an action.

Teams can keep their trackers and start with one workstream. Helen’s Foundry works best when teams ratify material premises and connect execution to them.

## Agent interface

Helen’s Foundry exposes the artifact graph through Model Context Protocol (MCP). Agents can:

- retrieve a bounded context chain before implementation;
- find implementation-ready work and claim the exact artifact;
- search linked product and strategy context;
- create and update artifacts through audited tools;
- connect work to the personas and problems it serves;
- inspect findings, including genuine conflicts and stale assumptions, before shipping.

MCP is the supported automation surface. The product also connects with GitHub and agent workflows including Claude, Codex, Cursor, and other MCP clients.

## Pull request alignment

The GitHub integration checks whether a pull request serves linked, ratified intent. It evaluates the diff against the artifact chain, reports acceptance-criteria coverage, and surfaces drift or missing context before merge. The goal is to answer whether the code is the right work, not only whether the code passes.

## Strategic coherence checks

Helen’s Foundry checks the graph using structural rules, recency analysis, semantic judgment, embedding geometry, property checks, code-symbol analysis, and portfolio-level signals.

- **Stale Premise:** Active work still builds on a parent that was abandoned, made obsolete, or marked a duplicate — a live effort resting on a discontinued foundation.
- **Stranded By Closure:** Open work left under a parent that already finished — every parent above it closed as completed or migrated, so the item needs a close, a new home, or an explicit carry-forward.
- **Coordination Collision:** Two active items under the same parent are being worked by different people at the same time, risking duplicated effort or a clash.
- **Lifecycle Contradiction:** An item's status fields disagree with each other — closed with no reason, a closure reason on something still open, or a priority where one doesn't belong.
- **Duplicate Singleton:** More than one live instance of something that should exist only once — like two missions, visions, or strategies — competing to be the source of truth.
- **Priority Inversion:** A child item is ranked more urgent than the parent it serves — high-priority work hanging off something that's been deprioritized.
- **Owner Over-Allocation:** One person owns more in-flight work than they can reasonably move at once — a context-switching and throughput risk.
- **Coverage Gap:** An active persona or problem has no live project, requirement, story, or descendant execution connected through SERVES or ADDRESSES.
- **Execution Context Gap:** Live stories that share a project or requirement declaration point cannot be traced to both an active persona and an active problem. The detector groups the repair at that declaration point while retaining exact per-story evidence; infrastructure-only tickets remain exempt.
- **Unowned Work In Flight:** Work that is actively in progress or in review has no owner — something is moving with nobody accountable for it.
- **No Prime Goal:** The workspace has several top-level objectives with no shared apex above them, so the work below them can ladder up to competing directions.
- **Decomposition Gap:** A mission, vision, strategy, or goal whose entire subtree never reaches doable work — intent that was stated but never broken down.
- **Priority Vacuum:** Work that's actively in flight has no priority set, so it can't be ranked against everything else competing for the same capacity.
- **Missing Accountable Owner:** A committed strategy or goal has no owner — a directional commitment with nobody accountable for whether it lands.
- **Resource Misallocation:** An objective carries urgent work underneath it that nobody is staffed on — a high-priority outcome that's effectively understaffed.
- **Zombie Project:** An active project whose work is all finished and that hasn't been touched in a long time — consuming attention with no live momentum.
- **Authority or Validation Needs Attention:** A committed premise lacks current human ratification, or its independent validation is missing or past its expected window.
- **Plan-of-Record Staleness:** A strategy or goal whose own description has gone stale while the work beneath it keeps changing — the plan frozen against what's actually happening.
- **Semantic Contradiction:** Two related items pull in logically opposite directions — an AI reviewer judged their intents to be in conflict, not just differently worded.
- **Redundant Overlap:** Two items in different parts of the graph say nearly the same thing — the same intent stated twice, eroding a single source of truth.
- **Chain Contradiction:** A child's intent opposes the parent it descends from — strategic drift along the line from objective to execution.
- **Copy-Paste Cascade:** A child simply restates its parent rather than breaking it into real work — an objective copied down the chain instead of decomposed.
- **Root Alignment Outlier:** A piece of execution work sits among the least connected, by embedding similarity, to its mission-root containment ancestor. SERVES/ADDRESSES completeness is measured separately by the structural demand checks.
- **Horizontal Incoherence:** An objective's children scatter in meaning rather than cohering as one decomposition — vertically aligned but pulling apart sideways.
- **Presupposed Capability:** A plan assumes a capability the workspace has no artifact for — the work presupposes something that doesn't yet exist.
- **Orphaned Metric:** A goal states a measurable target that none of its work actually moves — a metric nobody is working toward.
- **Portfolio Distribution Anomaly:** The work under a parent is lopsided along one dimension — nearly all critical, one priority flat, or one branch carrying the bulk — so that dimension stops distinguishing what actually matters.
- **Unrealized Commitment:** A committed piece of work has no resembling code in the connected repository — the intent is stated but nothing in the codebase looks like it.
- **Stale Implementation:** Work whose intent changed after the code that implements it last did — the spec moved on and the code hasn't caught up.
- **Signature–Doc Drift:** A public symbol whose doc-comment no longer matches its signature — the description and the contract have drifted apart.
- **Duplicate Implementation:** Two public symbols in different files are near-identical — duplicated logic or a missing shared abstraction.
- **Specification Clause Violation:** An accepted blocking clause with a reviewed exact checker has complete deterministic evidence that the pull request violates its required contract.

### Detector explainers

- [Browse all detector explainers](https://next.helensfoundry.com/detectors)

- [Stale Premise detector](https://next.helensfoundry.com/detectors/stale-premise)
- [Chain Contradiction detector](https://next.helensfoundry.com/detectors/chain-contradiction)
- [Execution Context Gap detector](https://next.helensfoundry.com/detectors/execution-context-gap)
- [Specification Clause Violation detector](https://next.helensfoundry.com/detectors/spec-clause-violation)
- [Stale Implementation detector](https://next.helensfoundry.com/detectors/stale-implementation)
- [Duplicate Implementation detector](https://next.helensfoundry.com/detectors/duplicate-implementation)

## Engineering leadership questions

- [Browse all engineering questions](https://next.helensfoundry.com/questions)

- [How do I verify AI-generated PRs against product requirements?](https://next.helensfoundry.com/questions/how-do-i-verify-ai-generated-prs-against-product-requirements)
- [How do I stop AI coding agents from using stale requirements?](https://next.helensfoundry.com/questions/how-do-i-stop-ai-coding-agents-from-using-stale-requirements)
- [How do I give AI coding agents the right product context?](https://next.helensfoundry.com/questions/how-do-i-give-ai-coding-agents-the-right-product-context)
- [How do I trace an AI-generated PR back to product strategy?](https://next.helensfoundry.com/questions/how-do-i-trace-an-ai-generated-pr-back-to-product-strategy)
- [How do I review more AI-generated PRs without slowing delivery?](https://next.helensfoundry.com/questions/how-do-i-review-more-ai-generated-prs-without-slowing-delivery)
- [How do I know whether an AI agent built the right thing?](https://next.helensfoundry.com/questions/how-do-i-know-whether-an-ai-agent-built-the-right-thing)
- [How do I prevent multiple AI agents from building the same feature?](https://next.helensfoundry.com/questions/how-do-i-prevent-multiple-ai-agents-from-building-the-same-feature)
- [How do I catch contradictions between tickets and strategy?](https://next.helensfoundry.com/questions/how-do-i-catch-contradictions-between-tickets-and-strategy)
- [How do I make product requirements machine-readable for AI agents?](https://next.helensfoundry.com/questions/how-do-i-make-product-requirements-machine-readable-for-ai-agents)
- [How do I audit why an AI-generated code change was made?](https://next.helensfoundry.com/questions/how-do-i-audit-why-an-ai-generated-code-change-was-made)

## Helen

Helen is the AI inside Helen’s Foundry. She acts as an editor for reasoning rather than prose: she reads the linked graph, asks questions when decisions conflict, and records the reasoning behind reconciliations.

## Integrations

GitHub, Anthropic, Linear, Notion, Figma, Slack, Discord, Sentry, and any MCP-compatible client.

## Access and contact

Helen’s Foundry is currently invite-only early access. It is free for early design partners for a limited time, and may have rough edges.

- Join the waitlist: https://next.hf.dev/waitlist
- Email: hello@helensfoundry.com
- GitHub: https://github.com/helensfoundry
- Status: https://status.hf.dev/
- Privacy: https://next.helensfoundry.com/privacy
- Terms: https://next.helensfoundry.com/terms