High Signal: Data Science | Career | AI Episode 23: Why Most AI Agents Fail (and What It Takes to Reach Production)
22 snips
Sep 2, 2025 Anu Bharadwaj, the President of Atlassian, shares insights on how AI agents are transforming teamwork across industries like publishing and finance. She discusses the success of bottom-up experimentation in creating effective AI tools and how non-technical teams are leading the charge. The conversation covers the importance of reliability in AI outputs, the future of proactive workflows, and the challenges of user interaction as AI evolves from tools to teammates. Anu also explores the exciting possibilities of multiplayer interactions between humans and AI.
AI Snips
Chapters
Transcript
Episode notes
Playbook To Move From Sandbox To Prod
- Anchor rollouts to a clear problem and measurable outcomes, iterate with real user feedback, and plan adoption/discovery.
- Use synthetic tests and broad user sampling to evaluate agent quality before full production.
Ground Agents With Sources And Reasoning
- Ground agents to vetted knowledge sources and enforce permissioned data connectors to reduce hallucinations.
- Surface reasoning traces and logs so humans can audit decisions and correct errors.
Self-Healing Is A Spectrum
- Self-healing agent behaviors are an extension of prompt engineering and built-in checks.
- Orchestration and trade-offs (cost, latency, accuracy) determine how much self-correction is practical today.