Invarra
We build control for AI that can act.
Invarra builds Phalanx, an execution-control gateway for AI agents. Our focus is the moment an agent's proposed action reaches a business system, where permissions, state, limits, and outcomes need to be explicit.
Useful autonomy needs a boundary you can operate.
Agents become more useful when they can do work across tools. Giving them that reach also creates a practical question: how do you keep their actions within the authority of the task?
Phalanx puts a controlled gateway at that point. The customer defines the task and rules. Protected connectors hold the credentials. The execution record makes the outcome available for inspection.
What we build around.
Authority you define
Execution stays within structured, authenticated task limits supplied by the customer.
Decisions you can inspect
Authorization follows explicit rules, and the action record preserves what was decided and what is known about execution.
Claims you can evaluate
Product claims stay tied to the protected workflow, configuration, evidence, and limitations that support them.
Founder
Sergio Valencia
Sergio leads Invarra's product development and research. His work connects questions about how AI systems behave with the practical controls needed to use them in consequential workflows.
Phalanx evaluations begin with a direct conversation about the system you are building and the action you need to control.
Research remains part of the work.
Invarra publishes research on semantic measurement and behavior under changes in representation. The Latent Invariance Principle and Canonical Semantic Realization address how to evaluate systems when meaning matters beyond surface wording.
Phalanx V1's execution controls operate without requiring these research methods or Phalanx-trained models.
Show us the workflow you want to make possible.
A specific action and a clear boundary are enough to begin the conversation.