This guide walks through the internals of the Beakr agent engine: how a chat request becomes a streaming graph execution, how agents hand work off to each other, how state is checkpointed, and how conversations stay within context windows even across long-running sessions. The engine is deterministic because it batches work into ordered super-steps, and resumable because every super-step produces a checkpoint that long workflows can pause, retry, or resume from.
In this guide
Pregel executor
Bulk Synchronous Parallel super-step loop -- plan, execute, update, checkpoint, repeat.
ReAct graph
The compiled graph: prepare, call_llm, route, execute_tools, handoff.
Agent catalog
Four stateless configs: researcher, analyst, coder, reviewer.
Meta-tools & handoff
delegate_to_agent, war_room, agent_pipeline, debate_ensemble.
Tool system
Registry, policy engine, context injection, parallel dispatch.
Checkpointing
Two levels -- super-step for ask_user, thread for cross-run continuity.
Compression
Four-layer working memory -- strip, surgical clear, summarize.
Key design decisions
Pregel / BSP model
Deterministic execution with built-in parallelism and checkpointing. Super-steps make every run reproducible and resumable.
Stateless agent configs
Agents are data (agent configs), not class hierarchies. Composable, swappable, testable.
Four-layer compression
Avoid expensive LLM summarization until absolutely necessary -- strip images, surgically clear compactable tool results, then summarize as last resort.
DB-writing tools serialize
Database-writing operations run one at a time for safety. Read-only operations run in parallel for speed.
Native streaming
The model's response streams to the frontend token-by-token with zero buffering -- the frontend sees tokens as the model produces them.
Two checkpoint levels
Mid-run checkpoints let the agent pause for user input and resume exactly where it left off. Thread-level checkpoints preserve conversation history across sessions.