Everyone can build an agent that passes a 5-step demo. But 35 steps into refactoring a real codebase, non-determinism catches up with you: tool calls fail, local state gets corrupted, and the agent burns thousands of tokens trying to patch code built on a broken premise. Standard retry loops don't fix this, they just compound context pollution until the run collapses.

In this talk, we stop treating agent execution like a simple prompt loop and start treating it like a database engine. We'll dive into the architecture of Transactional Agentic Workflows, bringing atomic execution, AST-aware snapshotting, and deterministic git/context rollbacks to long-horizon AI systems.

We will unpack:

Atomic Execution Units: Wrapping multi-file agent edits into rollback-safe, isolated transactions.

AST-Level Checkpointing: Using tree-sitter state graphs to checkpoint clean code states before an agent branches into complex refactoring.

Context Pruning & Rewind Mechanics: How to strip failed execution paths from prompt history without losing the diagnostic error signature.

Orchestrator-Worker Circuit Breakers: Detecting infinite fix loops early and triggering hard state rewinds before workspace corruption occurs.

You'll leave with a battle-tested architectural framework for building long-horizon agents that can recover gracefully from catastrophic failures in real-world, messy codebases.