v2.1.0
BREAKINGJanuary 2026Multi-Agent Systems & Project Association
- Multi-agent systems with sequential, parallel, and loop execution
- Session-based conversation management
- Project-scoped deployments (breaking change)
- Tools now optional for simple LLM agents
Added
- •Multi-agent support: agent_type with sequential, parallel, and loop modes
- •Sub-agents deployment with sub_agents field for complex workflows
- •Session management with session_id parameter and with_session() method
- •Planning & reasoning modes (planning_mode: builtin or plan_react)
- •Code execution support (code_execution with gemini, vertex_sandbox, or local executor)
- •Advanced LLM control (temperature, top_p, top_k, max_output_tokens)
- •Project-scoped API keys and agent deployment with project_id
- •max_iterations parameter for loop agents
- •thinking_budget for planning mode token allocation
- •include_thoughts flag to expose reasoning in responses
Changed
- •BREAKING: project_id is now REQUIRED in AgentConfig
- •BREAKING: display_name is now REQUIRED (no longer optional)
- •BREAKING: version is now REQUIRED (no default value)
- •BREAKING: timeout is now REQUIRED (no default value)
- •Tools are now OPTIONAL - empty list allowed for simple LLM agents
- •SDK always sends tools, memory_mb, timeout fields regardless of values
- •Backend validates project_id matches API key scope
Improved
- •Better validation error messages with detailed field information
- •Frontend deployment wizard allows agents without tools
- •SDK payload construction optimized for required backend fields
- •Documentation updated with comprehensive multi-agent examples