ACD

What the company appears to be: ACD is a self-reported project that describes itself as an AI orchestration system designed to maintain persistent intent across long-running tasks, model swaps, and context loss. It builds on the idea of an "intent graph" — a durable structure for managing AI agent collaboration.

What changed: The author states they built a system where AI agents coordinate through a graph-based structure that inherits intent hierarchically (current, secondary, major), records decisions and evidence in-band, and enforces safety constraints. This was demonstrated by porting a real open-source OS (AROS) to run native Chromium.

Single most important open question: Is there any evidence of commercial traction or product-market fit beyond this hackathon submission? The description contains no data on revenue, customers, usage, or adoption — only claims and self-reported technical execution.

The description states that ACD is not a chat loop with tools bolted on. Instead, it's a system where every action is represented as a node in a graph with three connections:

• Up — to an inherited intent hierarchy (current, secondary, major levels).
• Sideways — to decision records (AI_DECISION_CHAIN) that constrain the node.
• Down — to evidence: a manifest with content hashes registered at capture time.

The system also enforces invariants such as:

• Safety constraints apply to all instruction sources including human chat messages.
• Only an audited console can override safety guards.
• Failures are kept as signal, not deleted.
• Observability is treated as state, not history — meaning evidence must be captured live and cannot be retroactively added.

This system was used to coordinate three AI models (Codex, Claude, Antigravity) through a self-built relay (MCP server + web console), with minimal human involvement beyond setting top-level intent and minting time-boxed overrides.

The author makes several claims about what ACD does:

• It addresses structural failure in AI agents due to present-bias.
• It enables heterogeneous AI teams to hold one intent across days, hand-offs, and compression without human supervision.
• The core innovation is the "intent graph" — a durable, inheritable object that makes why first-class.
• It allows for autonomy that is legible and accountable rather than black-box.

These claims evolve from a critique of current AI agent behavior ("they spent so much time optimizing for efficiency they built in failure") to a proposed solution (the intent graph) and then to a demonstration (porting AROS to run Chromium).

The positioning seems to be: AI orchestration that stays on goal with minimal human steering, using an intent graph as the foundational structure.

Not evidenced. The description does not identify any specific customer segments or personas, nor does it describe how the system would be used in practice beyond this hackathon project.

Not evidenced. There is no mention of pricing models, monetization strategies, or business structure. The author only describes building a technical prototype and does not reference any revenue streams or customer acquisition efforts.

The description states that the system was built using:

• Three AI models: Codex (GPT-class), Claude, and Antigravity.
• A self-built relay: MCP server plus a web console.
• The team ran with minimal human involvement — setting top-level intent and minting time-boxed overrides.

It also reports:

• 68 jobs, 112 registered evidence records, 4,208 coordination events.
• Demonstrated work on real open-source infrastructure (AROS) including kernel/runtime fixes, API translations, and a working native Chromium port.
• Evidence was captured live with checksums to ensure bit-exactness.
• The system enforces observability as a hard gate — completion is refused unless hashed evidence manifests exist.

This suggests a high level of technical sophistication in both AI coordination and software engineering practices.

Not evidenced. There is no mention of users, customers, revenue, or adoption metrics beyond the hackathon submission. The description focuses entirely on what was built, not how it was used or received post-hackathon.

Not evidenced. No information is provided about competitors, market positioning, or competitive landscape. The author does not reference existing tools or platforms in this space.

• Unverified claims: All evidence comes from a single self-reported source — no independent verification.
• No traction data: No indication of product-market fit, user feedback, or commercial viability.
• Limited scope: The project is described as a hackathon submission with no clear path to production use.
• Highly technical and niche: The intent graph concept may be difficult for general audiences to understand or adopt.
• Founder-only team: Only one member (Timothy Deters) is listed, which raises questions about scalability and execution capability.

1. What specific problems in AI agent collaboration does ACD solve that existing tools don’t?
2. How would the intent graph scale beyond a single hackathon project?
3. Has there been any external testing or feedback on the system’s performance?
4. Are there plans to commercialize this technology, and if so, how?
5. What are the key assumptions underlying the intent graph model, and how might they break down in real-world usage?