The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Agent Swarms and Knowledge Graphs for Autonomous Software Development with Siddhant Pardeshi - #763
252 snips
Mar 10, 2026 
Siddhant Pardeshi, co-founder and CTO of Blitzy, builds autonomous systems that generate production-ready code at enterprise scale. He discusses hybrid graph-plus-vector maps for navigating large repos. Short sentences cover agent swarms and orchestration, dynamic agent personas and tool selection, why flat memories fail, and real-world evals and verification for production readiness.
AI Snips
Chapters
Transcript
Episode notes
Separate Context And Agent Engineering
- Solve autonomy by splitting context engineering and agent engineering to give agents the right info at the right time.
- Blitzy claims effectively infinite context via careful context compaction, dynamic agents, and orchestration techniques.
Use Hybrid Graph Plus Grep For Large Repos
- Combine semantic (vector/graph) search to narrow scope and grep to find exact lines to avoid missing usages in large repos.
- Use a hybrid graph-plus-vector map to jump instantly across millions of lines instead of burning tokens.
Orchestrate Swarms Not One Leader
- Recruit dynamic swarms of agents and use the database as orchestration to hyperscale tasks instead of a single leader orchestrator.
- Break specs into tasks recursively so thousands of agents can run in parallel without overloading one leader's context.