Pendulum: Human Attention Manager

The description states that Pendulum is an always-running attention agent for coding agents already running in tmux panes. It does not spawn or own workers; instead, it watches them and guards the user by implementing a mechanism called the "entropy gate." This gate uses a model-reported number (entropy: 0.0–1.0) and a human-set tolerance level (/gate low / medium / high). When entropy rises above the gate, actions are halted until the human answers queued questions.

It also supports features like:

• A /away command that starts negotiation before leaving.
• Pre-authorizations stored as standing instructions.
• A Telegram bridge for mobile access.
• An append-only ledger to track all actions and decisions.
• Deterministic test suite and demo environment.

The system is built using:

• Codex CLI (for building)
• GPT-5.6 Luna tokens via Pi SDK
• Node.js with TypeScript, SQLite, and custom tools injected into the model
• tmux captures for observation

Inference The product appears to be a proof-of-concept or prototype built by one individual in a hackathon context.

The author claims that human attention is now the bottleneck in multi-agent coding workflows, not agent capability. They state that only humans can reduce entropy in a codebase and that Pendulum is the missing piece to manage this scarcity.

They also claim:

• The system works end-to-end.
• It trusts the model’s self-reported entropy completely but logs all actions for auditability.
• It avoids mechanical rules by relying on prompts rather than code.
• The entropy rubric lives in one prompt file and improved more quickly than any code written.

Inference The positioning is that Pendulum addresses a real problem (human attention scarcity) in a novel way, using LLMs as both builder and executor. However, the claim of “end-to-end” functionality is not independently verified.

The description does not name specific customers or target industries beyond developers working with coding agents. It implies that users are individuals or teams who run multiple agents in parallel (e.g., tmux panes) and need help managing attention.

Inference The ICP likely includes solo developers or small teams using AI coding tools like Codex, where human attention is a limiting factor.

There is no evidence of pricing, monetization strategy, or business model in the description. The project was submitted as part of a hackathon and is described as self-hosted.

Not evidenced

The system uses:

• Codex CLI for building
• GPT-5.6 Luna tokens via Pi SDK
• Node.js with TypeScript, SQLite
• tmux captures for observation
• Telegram bridge for mobile access
• Deterministic test suite (37 tests)
• Append-only event ledger enforced by triggers

It implements:

• Entropy gate logic
• Pre-authorizations and standing instructions
• Auditability through ledgered actions
• Model hallucination fallbacks

Inference The technical stack is minimal but functional for a prototype. The system is designed to be deterministic, auditable, and self-contained.

The description contains no evidence of traction or adoption beyond its own demo. It was built in a hackathon setting and is described as self-hosted.

Not evidenced

There is no mention of competitors or competitive landscape in the description. The author does not reference existing tools for managing attention or agent orchestration.

Not evidenced

• Single-person team: No evidence of a larger team, which may limit scalability or long-term development.
• Self-reported only: All claims are unverified; no third-party validation or data on usage.
• Hackathon prototype: The product is presented as a hackathon submission, not a mature product.
• No revenue or customer data: No evidence of monetization, users, or real-world impact.
• Model reliance without external checks: Trust in model behavior without independent verification.

1. What is the actual level of human involvement required to operate Pendulum? Is it truly autonomous?
2. How does Pendulum handle cases where models hallucinate or misreport entropy?
3. Has the system been tested beyond the demo environment, and in what conditions?
4. Are there any known limitations or edge cases that were not addressed in the demo?
5. What are the plans for scaling beyond a single-user, self-hosted model?
6. How does Pendulum integrate with other agent platforms or CI/CD systems?