Skip to main content
Zeus’s Agent mode is built on the DeepAgents framework, supporting autonomous task planning, tool invocation, and execution. The runtime manages the Agent’s lifecycle — from workspace initialization, tool assembly, and prompt construction to session isolation and state persistence.

Core Modules


Workspace

Each user has an independent cloud workspace managed by CloudDriveBackend, backed by Supabase Storage for persistence and Redis for caching.
PathDescription
users/{user_id}/workspace/Agent’s working directory for output files
users/{user_id}/workspace/projects/Project files
users/{user_id}/workspace/sandbox-output/Sandbox execution results
users/{user_id}/workspace/uploads/User-uploaded files
users/{user_id}/memory/Long-term memory files

File System Detailed Design

Learn about the complete design of CloudDriveBackend, Checkpoint, and other storage architecture

Sessions

Each conversation creates an independent Session for state isolation:
  • Session ID: Format is session_{hex12}, provided by the frontend or auto-generated
  • Thread ID: Same as Session ID, used for Checkpointer state isolation
  • Context Cache: Stores tool list, system prompt, and interrupt configuration per session_id, reused during HITL recovery

Checkpointer

State persistence is implemented via LangGraph’s Checkpointer mechanism:
EnvironmentImplementationDescription
ProductionPostgresSaverPostgreSQL persistence, supports cross-process recovery
Development/FallbackMemorySaverIn-memory storage, lost on process restart
The Checkpointer automatically saves the complete state after each Agent call (messages, tool calls, Agent internal state), enabling recovery after HITL interrupts and page refreshes.

Checkpoint Storage

Learn about PostgresSaver detailed configuration and HITL recovery flow

Modes

Zeus supports three interaction modes, each constraining the Agent’s tool set and behavioral boundaries:

Mode Comparison

FeatureAgentAskPlan
File readYesYesYes
File writeYesNoNo
Sandbox executionYesNoNo
Memory readYesNoYes
Memory writeYesNoNo
HITL approvalYesNoNo
Tool callsAllRead-only subsetRead-only subset
Typical scenariosCoding, deployment, automationCode explanation, Q&AArchitecture design, plan comparison
Modes take effect immediately at the invoke() entry point, implemented through disable flags and tool filtering. Ask mode disables sandbox, memory writes, and HITL; Plan mode disables sandbox and HITL but retains memory read access for context.