Building AI Agents for Enterprise IT Automation: Thira's Approach to Trust and Learning
Thira is developing an agentic "system of execution" for enterprise back-office IT processes, leveraging AI to automate complex workflows across disparate systems. The core system design challenge involves building a self-learning engine that can adapt to unique enterprise environments, while simultaneously ensuring trust through features like audit trails, kill switches, and semi-autonomous modes.
Read original on The New StackThira, founded by Apptio veterans, aims to automate repetitive IT processes within enterprises using AI agents. Unlike traditional systems of record, Thira is designed as a "system of execution," focusing on acting upon data rather than just measuring it. This involves integrating with and extending existing enterprise systems like ServiceNow and Workday.
Architectural Pillars: Discovery, Learning, and Trust
At its heart, Thira's platform relies on a discovery and self-learning system. Agents ingest diverse data sources (documents, slide decks, diagrams) to learn how a particular enterprise environment is structured. This learning culminates in dynamic "execution maps" that define deterministic paths for autonomous operations and flag non-deterministic ones requiring human judgment.
The Trust Problem in AI Agents
When AI agents are tasked with interacting with critical IT systems, trust becomes paramount. Thira addresses this by incorporating kill switches, comprehensive audit trails, and rollback capabilities. It also offers a semi-autonomous mode where humans can credential actions, providing a transitional phase for enterprises to gain comfort before moving to fully autonomous operations. This layered approach to trust is a critical design consideration for any system interacting with sensitive enterprise data.
Overcoming Integration Complexity
A significant technical challenge for Thira is teaching agents to operate across the highly varied and often inconsistently configured systems within different enterprises. The solution proposed is a flywheel effect: as more customers use the system, the learning engine improves by encountering and adapting to diverse environments, ultimately benefiting all users. This approach suggests a shared learning model or federated learning concept.
- Discovery & Mapping: Agents ingest unstructured and structured data to build execution maps.
- Execution Logic: Distinguishing between deterministic (agent-driven) and non-deterministic (human-assisted) workflows.
- Trust Mechanisms: Kill switches, audit trails, rollbacks, and semi-autonomous modes.
- Learning Flywheel: Continuous improvement of agent intelligence through exposure to diverse customer environments.
The focus on integrating with over 40 disparate systems, which are often manually managed or scripted today, highlights the complexity of enterprise IT automation. Thira's strategy is to abstract this complexity through its learning engine, creating a more cohesive "system of execution" atop fragmented existing infrastructure.