MLA: Live Decision Coordination for Codex

The description states that MLA is a live decision coordination layer for humans and AI agents. It monitors decisions as work happens across sessions and team members, flags conflicts, helps agents revise their approach, and escalates to people when human judgment is required.

It integrates with OpenAI Codex using three capabilities:

• Prompt-time governance: Injects instructions into Codex prompts to guide how project evidence should be treated.
• Governed retrieval: Uses MCP tools to retrieve current decisions, superseded decisions, constraints, evidence, and citations.
• Action interception: Evaluates proposed actions before execution against governed rules, returning warnings or denials.

The system is built using:

• CLI
• Codex
• GPT-5.6
• Hooks
• MCP (Model Control Protocol)
• PostgreSQL
• Python
• TypeScript

It operates via two hook events in Codex:

• UserPromptSubmit: Invokes MLA’s prompt-grounding flow.
• PreToolUse: Evaluates tool use against existing rules.

MLA also includes an MCP integration that allows Codex to retrieve governed information such as current decisions, superseded ones, constraints, and citations. Installation is deterministic and idempotent, preserving user-owned hooks.

The author states that:

• Coding agents are becoming active participants in software delivery but lack a shared understanding of team decisions.
• Traditional search and RAG do not resolve whether information is current or superseded.
• MLA addresses this by acting as a live decision coordination layer for humans and AI agents.

They claim that MLA:

• Keeps work consistent while multiple people and agents operate in parallel.
• Identifies when new work conflicts with existing decisions.
• Gives agents context to revise their approach without waiting for human input.
• Surfaces conflicts requiring human judgment instead of silently choosing a side.
• Provides governance through prompt-time instructions, governed retrieval, and action interception.

These claims reflect an evolution from general AI agent coordination toward real-time decision management within specific tools like Codex. The positioning is not yet market-tested or validated beyond the demo context.

The description does not clearly identify a target customer segment or ideal customer profile (ICP). It implies that MLA targets:

• Developers working with AI agents (e.g., Codex),
• Teams managing multiple concurrent development sessions,
• Organizations using structured decision-making practices in software delivery.

However, no explicit mention of enterprise customers, developer tooling companies, or specific use cases beyond the demo is provided. The system appears to be aimed at developer teams using AI-assisted coding tools, but there is no evidence of segmentation or targeting strategy.

There is no evidence in the description of a business model or pricing structure. The project is presented as a hackathon submission for OpenAI Build Week and lacks any indication of monetization, licensing, or customer acquisition strategies.

The system:

• Reuses existing MLA components including governance core, rule evaluator, enforcement audit path, and MCP server.
• Integrates with Codex via two hook events (UserPromptSubmit, PreToolUse).
• Uses a thin wrapper for prompt grounding and directly reuses PreToolUse enforcement logic.
• Bundles its MCP server as a native Codex plugin.
• Supports deterministic and idempotent installation that preserves user hooks.
• Requires explicit trust of hooks via /hooks before enforcement activates.

The author notes:

• The system distinguishes between retrieval and authority.
• It tracks decision status (current/superseded) separately from documents.
• Enforcement is warning-first by default; hard blocks are enabled only for specific enforceable rules.
• The connector avoids duplicating governance logic, reusing existing systems.

This suggests a modular, reusable architecture with clear separation of concerns and minimal duplication.

There is no evidence of traction or maturity beyond the demo. The project:

• Was submitted to a hackathon.
• Has no stated revenue, customers, or adoption metrics.
• Does not reference any prior versions or production deployments.
• Is described as an extension of an existing system, but without details on its prior usage.

The author mentions that the system was tested in a limited development workflow and connected to a benchmark (stale-context research), but no broader deployment or validation is reported.

There is no evidence provided about competitive landscape or direct competitors. The description does not name other tools or platforms offering similar decision coordination for AI agents or developers. It focuses solely on the novelty of integrating live decision management into Codex, without comparing it to existing solutions in the space.

• No commercial traction: No evidence of revenue, customers, or adoption beyond a demo.
• Limited scope and testing: The system is demonstrated only within a narrow development workflow and tested Codex version.
• Dependency on user trust: Enforcement only works if users explicitly trust hooks in Codex — otherwise, governance remains inactive.
• Not production-ready: The project is described as a hackathon effort with no indication of scalability or robustness beyond the demo.
• Unclear value proposition for end-users: While it addresses coordination issues, there’s no clarity on how this translates into measurable benefits for teams or organizations.

1. What is the existing MLA system used for outside of this Codex integration?
2. Has the system been tested in real-world development environments with multiple users and agents?
3. How does MLA handle conflicts that arise from decisions made by non-human actors (e.g., automated systems)?
4. What are the limitations of the current implementation regarding scalability or performance?
5. Are there plans to support other AI coding platforms beyond Codex?
6. How is decision provenance tracked and audited in practice?
7. What kind of feedback has been received from developers using this system?
8. Is there a roadmap for moving beyond the demo phase into production use?