Today, we’re announcing Fulcrum Research, a startup scaling human oversight. We are building debuggers that tell you why your agents fail, and what your rewards are truly testing for—the first step toward the inference-time infrastructure required to safely deploy agents.

Check it out on GitHub! https://t.co/CRxg4LuwFu

Introducing Quibbler MCP! Now you can use Quibbler to critique any coding agent, not just Claude Code. Here's a demo of Quibbler working on Cursor:

.@fulcrumML, one of the highest firepower s25 teams I worked with is open-sourcing some practical tools for coding agents today: orchestration, monitoring, and critique. The best AI tools don't just make models more capable, they make people more capable. Orchestra from the

Real-time oversight of agent systems will be critical as agents scale. If you’re interested, we’d love to hear from you. We’re hiring!

We think oversight features, like the ones we built into Quibbler and Orchestra, are what makes it possible for parallelization to be useful, and not destructive.

In the worst case, your attention is spent on the behavior of your agents: preventing them from taking unsafe actions, making sure they're not lying to you, and understanding if their outputs are correct.

In the best case, coding with agents allows your attention to be spent on the “right” parts of your code: its functionality, architecture, and its failure-modes.

Orchestra enables true multi-agent coding: parallel execution, active coordination, and full visibility of your coding agents. A designer spawns executors working in isolated environments. When an executor needs help, it messages the designer (getting your input if needed).

Quibbler uses hooks to review what your agents are doing, making sure they’re running tests, following your coding style, and not fabricating results. In longer running tasks, we found Quibbler useful in enforcing intent, allowing us to check in on our agent less.

Try them out and install on PyPI: Quibbler: https://t.co/7t5BuKqjMb Orchestra: https://t.co/FEwolma1u2 Launch post:

live monitoring of agents in real-world settings allows us to learn lessons about AI control that can later be applied to x-risk mitigation. hoping this open-source launch helps control researchers experiment with more realistic threat models!

Today, we’re excited to announce that we’re open-sourcing two tools: Quibbler is a background agent that critiques your coding agent’s actions. Orchestra is a multi-agent coding system: it uses a designer agent that spawns and coordinates your parallel coding agents.

> Potemkin set up painted facades along the riverbank, so that [...] Catherine would see beautiful villages – each just a couple of inches thick. The rise of AI agents makes the Potemkin problem commonplace. Cool-looking scalable oversight & agent observability startup.

Or reach out @KaivuHariharan @uzpg_

If you are building RL environments, evals, or deploying agents, contact us to try our tooling. https://t.co/a6AN2CWSmj

We aim to solve the problem of overseeing agents and understanding their effects on the world. Check out our post on why it will be challenging: