Narrow permissions
Each agent gets a defined role and only the tools that role needs. Credentials stay out of the agent’s reach.
AI agents you can verify
What an agent should be permitted to do. How its work is checked before anyone relies on it. Whose interests it actually serves.
The question we are working on
Capable AI agents are arriving faster than the judgment about how to use them.
Most of the industry treats AI safety as a refusal problem: teach the system to say no, count the refusals, call it aligned. We think that gets it backwards. A filter that stops honest users while sophisticated actors route around it is not safety; it is the appearance of safety, and the appearance is worse than an honest account of what the system will actually do. The work that matters is verification, not suppression.
That argument is set out in full in Negative Alignment, and it is not theoretical for us. The same claim is built into how we run our own agent systems: enforced roles, adversarial review before anything is called done, and a record that outlives the session.
What it means for client work
Independent review
Work is checked by something that did not produce it, and against the original requirements, before it is called finished.
A person in charge
Anything consequential, such as spending money, changing a live system or sending a message outside the business, needs a human to approve it.
Questions about agents in your business? See the AI agent FAQ or discuss your project.
Essays and field reports
Giving AI Agents a Memory That Outlives the Session
Every agent session starts with amnesia. SGS runs an LLM-maintained wiki and a requirements ledger so the next session starts ahead. The hard part was not storage; it was keeping the memory honest.
Why We Make One AI Review Another's Work
Code written with one AI model is reviewed by a different vendor's model before SGS calls it done. What cross-model review has caught, where the reviewer is wrong, and three rules learned on real releases.
Meet Choir: Giving AI Work Direction and Accountability
Curious about AI for your business? SGS helps you find a practical starting point, backed by Choir's coordinated AI work and review.
Negative Alignment
Why current AI safety implementation is net harmful, and what the alternative looks like — a structural diagnosis of negative alignment (an asymmetric filter that lets sophisticated actors through while stopping honest users) and a three-layer sketch of the solution space.