Yigdesk

The description states that Yigdesk is:

• An adaptive decision workbench for agents and humans.
• A six-operation MCP blackboard, with operations including open_decision, propose_candidate, post_claim, read_board, cast_approval, and request_resolve.
• A system with a deterministic engine that prices candidates from the same immutable source.
• A read-only blackboard with an append-only ledger, where board state is a pure fold of JSONL operation logs.
• A trusted adaptive UI, using declarative manifests to control agent influence without allowing unsafe HTML or JS.
• A system that enforces a fail-closed evidence model, where claims are accepted only if all references resolve to real evidence.
• A human-action bridge with allow-listed envelopes for actions, including decision, candidate/scope, action type, and correlation ID.

The author also states that Yigdesk supports:

• Codex plugin, standalone browser, and ChatGPT Developer Mode components using the same core MCP server and ledger.
• A durable identity and cutoff mechanism to freeze input sequences after approval or resolution.
• A local workflow continuation adapter, where a ledger watcher resumes workflows and ChatGPT can request follow-ups.

Inference Yigdesk appears to be a conceptual framework for managing AI agent decisions in a deterministic, auditable, and human-governed way. It is not a commercial product but a prototype or proof-of-concept built for a hackathon.

The author states:

• Yigdesk is built for AI agents to see consequences and evidence before acting.
• The goal is not to make agents agree, but to make their disagreement comparable, grounded, and safe enough for a human to decide.
• It aims to solve the problem of different spreadsheets, assumption sets, and definitions of “correct” across functions like Finance, Sales, or Operations.
• It introduces a Codex-native place where agents can keep perspectives while being forced to calculate inside the same measurable environment.

Inference The positioning is that Yigdesk is a decision-making infrastructure for AI agents in enterprise settings. It is positioned as a tool for shared state transition systems, not just opinion exchange, and it emphasizes determinism, auditability, and human accountability.

The description states:

• The target is AI agents, particularly those working in enterprise environments like FP&A.
• It is built for Codex-native workflows, suggesting a focus on developers or enterprises using OpenAI’s Codex or similar tools.
• The author mentions that the system supports multi-agent collaboration across roles such as CFO, COO, and Commercial Director.

Inference The ICP appears to be enterprise developers or AI teams working with AI agents in structured decision-making contexts. It is not yet clear if there are specific customer segments beyond the hackathon demo.

Not evidenced.

The description states:

• Built with CSS, JavaScript, Python.
• Uses a six-operation MCP surface for agent interaction.
• Implements a model-from-data evaluator using exact decimal arithmetic and fixed rounding.
• Operates on an append-only blackboard, with JSONL operation ledger.
• Enforces fail-closed evidence — claims are accepted only if all references resolve.
• Uses a trusted adaptive UI with declarative manifests, no HTML/JS from agents.
• Implements a narrow human-action bridge with allow-listed envelopes and idempotency.
• Supports durable identity and cutoff, freezing input sequences after approval or resolution.
• Works across Codex, browser, and ChatGPT components using the same core.

Inference The technical architecture is designed for security, auditability, and deterministic behavior. It is built with a focus on state management, identity control, and human-in-the-loop decisioning, rather than scalability or performance for large-scale use.

Not evidenced.

Not evidenced.

• The project is described as a hackathon demo — no evidence of commercial traction, adoption, or real-world usage.
• There is no mention of revenue, customers, or funding rounds.
• It is built for Codex-native workflows, which may limit its applicability to broader AI agent ecosystems.
• The system is described as local, synthetic, auditable, and honest, suggesting it’s not production-ready.
• The author states that the public repository cannot claim a private native Codex callback, indicating limitations in integration or scalability.

Inference The project is in an early prototype phase. It lacks commercial evidence, and its design may be too narrow for broader adoption without further development.

1. What are the specific use cases or workflows where this system would be applied beyond the hackathon?
2. Is there any plan to integrate with production AI agents or enterprise systems beyond Codex and ChatGPT?
3. How does Yigdesk handle multi-tenant environments, and what are the plans for scaling to multiple users or organizations?
4. What is the roadmap for moving from a local, synthetic demo to a production-ready system?
5. Are there any partnerships or integrations with AI agent platforms or enterprise software vendors?