The Hard Truth About AI Agents
summary
Digits runs user-facing accounting agents by reimplementing the 100 lines of code core loop in Go, reusing internal RPCs (extracting comments to auto-generate JSON tool definitions) instead of copying Python frameworks. Observability uses an agent-tracking tool with OpenTelemetry and reveals token-wasting loops and failed sessions. Memory lives as an internal microservice to avoid vendor lock, guardrails run via a separate LLM for verification, and a simple task-planning step cuts recursive tool calls and overall latency; a demo exposes context loss from token limits.
AI-written summary · about
chapters (11)
- 0:00Introduction and speaker welcome
- 1:10What are AI agents really
- 4:43Agent use cases in production
- 7:11Building agent infrastructure
- 13:38Lessons learned from frameworks
- 16:08Why observability matters
- 17:58Live demo of memory challenges
- 21:10Task planning and guardrails
- 23:35Demo of agent failure analysis
- 25:43Key takeaways and conclusions
- 26:27Q&A on reward systems
related talks
01h 01mAI Agents In-Depth – Function Calling, MCP and Tool Use Under the Hood
47mValidation and Observability of AI Agents
19mYour Biggest User Isn't Human: Designing DevTools for AI Agents
30mAI Agents Unchained
40mBuilding Effective AI Agents Simple, Composable Patterns
34mAccountability in the Age of AI Agents
40mAI Agents in Practice
21mDemystifying AI Agents: A New Primitive for Flow Control