Research for the agentic enterprise. We tackle the hardest problems in enterprise AI. Controlling agents. Ensuring accountability. Orchestrating predictable outcomes.
// Featured research
Operationalizing Governed Agentic Workflows in Regulated Finance
Regulated industries such as finance and healthcare face critical governance challenges that hamper the adoption of AI-assisted software development and hinder the transition …
Mean estimated effort reduction Pricing execution platform
Mean estimated effort reduction Client deal negotiation platform
The Ten Commandments for Governed Agentic AI
Ten architectural principles for controlling autonomy, preserving accountability, and building trust in enterprise AI systems.
- Agentic AI shifts the governance challenge from controlling individual model outputs to controlling actions across interconnected systems.
- Incidents involving destructive operations, unauthorized access, and AI-enabled cyberattacks demonstrate that traditional model-level safeguards are insufficient when software can act independently.
// The agenda
Three questions we’re working to answer. Agents are moving into production faster than the evidence on how to govern and measure them. We believe organizations should have that evidence before agents carry real responsibility, and we are exploring how to produce it.
01 / Governance and provenance
How can agents act within clear authority and limits? Exploring control across agent ecosystems, the provenance of agent work, and the risks introduced by generated code.
- Cross-vendor agent governance
- Attributable, reviewable agent actions
- Code provenance and licensing risk
- Automated human escalation
02 / Applied research
What do we learn when agents meet production? Using our own implementation experience to examine enterprise agentic workflows, review and the operational realities of agentic work.
- Limits of automation and human intervention
- Agent-authored pull requests at scale
- Agents on high-value production assets
03 / Capability And evaluation
How should governed agentic work be measured? Proposed studies on intent engineering, software engineering benchmarks and agent memory, considering capability alongside governance.
- Intent quality and output
- Governance-aware software benchmarks
- Agent memory usefulness and reliability
// Our Research
An internal study of 49 enhancement requests across two production codebases, examining estimated effort savings, human intervention and governance controls.
Remaining Sovereign in an Agentic World
An internal study of 49 enhancement requests across two production codebases, examining estimated effort savings, human intervention and governance controls.
How the Hugging Face attack was allowed to happen
An internal study of 49 enhancement requests across two production codebases, examining estimated effort savings, human intervention and governance controls.
Agentic control and authority
Operationalizing Governed Agentic Workflows in Regulated Finance
// Upcoming Research
The next questions we want to explore. AI and agentic capabilities are expanding faster than any technology before. We track every shift and explore where expanding capability meets enterprise needs.
Human escalation – How can escalation decisions be evaluated against real banking regulation, so agents catch the cases that need a human without interrupting the ones that don’t?
Conversational personal finance security – How should conversational finance assistants be secured and evaluated against indirect prompt injection, when no model is immune and attack success varies widely between them?
Agent gym – How can realistic, sandboxed scenarios measure whether guardrails limit unsafe agent actions and whether agent behavior stays within regulatory compliance?
// From inquiry to implementation
Explore how this thinking informs AIOS. Our research asks what governed agentic work requires. AIOS is where we put those ideas into practice.





