wtf is omniagents?

OmniAgents is a framework for building AI agents that are designed to keep working on a task until the result is genuinely as good as it can get, rather than stopping at the first plausible answer. It works with multiple AI model providers, including Claude, GPT, Gemini, Grok, and local models. The core idea is that most agent systems stop too early, producing a single answer and summarizing it, while OmniAgents pushes the agent to use every available tool, skill, and approach before declaring the work done. The framework is built around six rules. An agent must not yield early: it stops only when acceptance criteria are met and no remaining tool or retry would improve the result. The agent must consider the full capability surface, meaning every relevant tool, installed skill, available model, and memory of past successes and failures. Each run produces a capability ledger that records what worked, what failed, and what would be needed to unlock the failed items. This ledger carries forward, so the system remembers its limits and improves over runs. Humans remain in control by setting goals, authority limits, kill switches, and budgets for tokens, money, and time. The system is provider-agnostic, so it is not tied to one vendor, and it operates natively in the operating system, terminal, browser, and filesystem rather than functioning as a chat widget. The architecture centers on an Effort Governor, which runs a loop of planning, acting, observing, recording in the ledger, and replanning. If the agent stalls, it escalates to a stronger model, a different tool, or a human. A Verification Gate requires proof that the deliverable works, such as test output or screenshots, before the agent can declare completion. The code example shows installation via Bun and a command-line interface for running agents against goals and inspecting the ledger afterward. The README also describes a broader ecosystem of related projects under the LoveLogic umbrella, including a multi-agent deliberation layer called Hermes Council and a governance layer called Control Plane. The repo map at the bottom is cut off. The full README is longer than what was shown.