Industry Research

Systems That Act: Decide, Execute, Verify, Improve

Operating the Physical World, Safely

I focus on the software layer that makes physical operations actually run: robotics, industrial automation, and any domain where software has to change what happens in the real world.

The core thesis is that the biggest winners will not be better systems that observe work (dashboards, analytics, copilots). They will be systems that can own execution safely. I call these Systems That Act.

In physical operations, value concentrates in the execution loop: take an action, measure the real-world outcome, and improve — under strict safety, uptime, and accountability constraints. The wedge I find most compelling is reliability-first: a RobOps version of DevOps / SRE. Start with fleet monitoring, remote ops, incident response, and exception routing with outcome tracking to earn trust. Then expand into permissioned autonomy, where software takes bounded actions within guardrails and human override. That is where compounding happens, because performance improves continuously across sites.

Until deployment and reliability are solved, robots, humanoids, and foundation models stay stuck at single-site experiments. This research is about the stack, the incentives, and the companies building the layer that gets you from 1 site to 100.

Overview

Framework for Execution

Systems That Act are platforms that:

  • Decide what is allowed under which rules
  • Act across people, software, and machines
  • Observe the outcome, detect problems, and recover
  • Learn by evaluating and improving the next time

The next generation of software platforms will not just explain what is happening. Winners shift from visibility to connecting safe decisions, execution, observability, and continuous learning across humans, robots, and software

Priorities Shift

  • AI can propose actions, but who ensures that they are good and safe
  • Automation is moving from pilots to production, where trust decides adoption
  • Advantage compounds through integration depth, verified outcomes, and operating history

Applications

  • [Focus] Physical operations and automation: robots, inspection, warehouses, where failure is expensive and measurable
  • Adjacent domains: enterprise agent workflows, infrastructure operations, security automation

Archetypes

Safety & Decision

Control of access and permissions, safe rollouts, undo, and audit trails

Execution & Orchestration

Constraints-driven decision-making & actions

Observability & Recovery

Detection and mitigation for systems that acted (up)

Learning & Improvement

Results check, silent failures catch, and feedback learning loop

Systems that Do Not (Usually) Act

Systems of record (ERP / CRM), dashboards, analytics, point tools, AI copilots without execution, heavily customizable software services

Stick figure poking AI with a stick saying C'mon, do something