feat: Agent Blocks for composable agent #11055

ogabrielluiz · 2025-12-17T12:40:40Z

Note

This is a Proof of Concept PR to validate the agent blocks architecture. Once the approach is established, this will be broken into smaller, more modular PRs for easier review. (This PR has ~9k new lines but most of them are tests 😅)

Summary

Introduces Agent Blocks - a set of composable primitives for building agent workflows visually:

AgentStep: The reasoning core - sends messages to LLM and routes based on tool calls
ExecuteTool: Executes tool calls from AgentStep with parallel execution support
WhileLoop: Loop control for iterative agent execution
AgentLoop: Pre-composed component that encapsulates the complete agent loop

Key Features

Conditional routing based on tool calls (AgentStep outputs to either ai_message or tool_calls)
Parallel tool execution with proper error handling
Streaming support with immediate tool notifications
Built-in session memory - AgentLoop auto-fetches chat history from Langflow's session memory
Full message history preservation via DataFrame
Think tool support for step-by-step reasoning

Architecture

ChatInput → WhileLoop → AgentStep → [Tool Calls] → ExecuteTool → WhileLoop
                              ↓ [AI Message - done]
                         ChatOutput

Memory Integration

AgentLoop automatically retrieves conversation history from Langflow's session memory:

message_history - Optional DataFrame input (takes precedence if provided)
n_messages - Number of messages to retrieve (default: 100)
context_id - Optional context ID for memory isolation

Benchmark Results (vs other agent frameworks)

50 iterations, 100% success rate, using gpt-4o-mini:

Framework	Version	Avg Init	Avg Runtime	Avg Total
Langflow	latest	0.89ms	1936ms	1937ms
LangGraph	1.0.5	12.33ms	1923ms	1935ms
Pydantic-AI	1.34.0	8.26ms	1941ms	1949ms
Agno	2.3.13	0.04ms	2934ms	2934ms

Runtime differences between Langflow, LangGraph, and Pydantic-AI are within API latency variance (~18ms). Langflow's AgentLoopComponent performs on par with the best agent frameworks.

Test plan

Unit tests for all agent block components
Integration tests for agent graph builder
Contract tests for agent loop events
E2E tests for tool calling workflows
Manual testing in UI

Add WhileLoop, AgentStep, ExecuteTool, and ThinkTool components: - WhileLoop: Flow control component with initial_state input for MessageHistory integration and loop feedback support - AgentStep: LLM reasoning step with conditional routing (ai_message or tool_calls output based on model response) - ExecuteTool: Executes tool calls with parallel execution and timeout support, outputs updated conversation as DataFrame - ThinkTool: Optional tool that lets the model reason step-by-step Supporting infrastructure: - LCModelComponent base class for model components - Message and tool execution utilities - Updated serve_app to use execute_graph_with_capture - EventManager improvements for streaming - Frontend cleanEdges fix for group_outputs components These components enable visual agent loops: ChatInput → WhileLoop → AgentStep → [tool_calls] → ExecuteTool → WhileLoop ↓ [ai_message] ChatOutput

Add Agent Blocks category to the frontend UI: - Add nodeColors entry with violet color (#7C3AED) - Add nodeColorsName mapping - Add SIDEBAR_CATEGORIES entry after Flow Control - Add categoryIcons with Blocks icon - Add nodeIconToDisplayIconMap entry

…sses

… and parent message handling

…utputs for conditional routing

…mportant messages based on vertex configuration

…is correctly assigned and all fields are included

… graph execution

… vertex identification

…nditional routing

…class

…odule

…of FakeCallModelComponent

…atOutput

…p functionality

Introduces comprehensive unit and contract tests validating the agent loop’s graph execution, LLM invocation, streaming behavior, and tool-call lifecycle. Ensures single AI message updates, immediate tool notification during streaming with a parent message, and reuse of tool entries across execution to avoid duplicates. Improves reliability and UX by formalizing event sequencing (accessing → executed → final), enforcing one message ID per response, and preventing duplicate message events. Supports performance by minimizing DB updates during streaming and tightening tool-content updates. Relates to feature X to enhance user experience and optimize performance.

- Check model_message.text instead of lf_message.text for unconsumed generator (lf_message loses generator reference after serialization via model_dump) - Remove redundant aggregation in fallback loop - aggregation already happens inside stream_and_capture() via nonlocal aggregated_chunk - Fixes corrupted tool_calls (e.g., 'calculatorcalculator') when running without event_manager

coderabbitai · 2025-12-17T12:40:47Z

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

✨ Finishing touches

🧪 Generate unit tests (beta)

Create PR with unit tests
Post copyable unit tests in a comment
Commit unit tests in branch feat/agent-blocks

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

…ext ID inputs

…rokenEdgesEdges Fixed handle reconstruction to match how NodeOutputParameter builds handles: - Regular outputs use [selectedType] - Loop outputs with allows_loop=true use [selectedType, ...loop_types] This fixes loop connections being incorrectly removed as invalid when loading flows with While Loop and Execute Tool components. Added comprehensive unit tests for both functions covering edge validation, loop edges, group_outputs handling, and hidden field filtering.

- Add SharedContextEventsDisplay component for rendering shared context updates - Add sharedContextStore for managing shared context state - Update ContentDisplay to include shared context events - Add SharedContextEventOutput type to chat types - Update buildUtils to handle shared context events

…Display component

…w utility functions

- Fix extract_loop_output to handle dict outputs from model_dump() - Add loop subgraph execution infrastructure - Preserve component configuration across subgraph deepcopy

ogabrielluiz added 21 commits December 15, 2025 21:39

feat(unified_models): implement lazy loading for individual model cla…

1e46c08

…sses

refactor(chat): improve model name extraction logic in ChatComponent

fd466fa

feat(agent-graph): implement agent graph builder with execution context

c61ea42

feat(agent-step): enhance streaming with immediate tool notifications…

dbca0f3

… and parent message handling

feat(component): enhance output processing logic to include grouped o…

ae34748

…utputs for conditional routing

feat(component): enhance message handling logic to prevent skipping i…

6e1069a

…mportant messages based on vertex configuration

feat(component): improve message event handling to ensure message ID …

e28394c

…is correctly assigned and all fields are included

feat(component): prevent overwriting explicitly set parameters during…

31dfd7f

… graph execution

feat(tests): add integration tests for agent graph builder module

c48735e

feat(component): update chat output/input checks to use base_name for…

565249f

… vertex identification

feat(tests): add async test for state model with group outputs and co…

e68cf54

…nditional routing

feat(graph): preserve start_component_id during preparation in Graph …

61c4f6e

…class

feat(tests): add unit and integration tests for agent graph builder m…

7d4d385

…odule

feat(tests): update test cases to use FakeAgentStepComponent instead …

6ed3aa4

…of FakeCallModelComponent

feat(chat): add clarification comments on message storage logic in Ch…

9cec18c

…atOutput

feat(agent): add AgentLoopComponent to encapsulate complete agent loo…

19d1d2e

…p functionality

feat(dependencies): add langchain-openai to integration dependencies

97422bc

github-actions bot added community Pull Request from an external contributor enhancement New feature or request and removed enhancement New feature or request labels Dec 17, 2025

ogabrielluiz removed the community Pull Request from an external contributor label Dec 17, 2025

feat(agent): enhance AgentLoopComponent with message history and cont…

0e9549e

…ext ID inputs

github-actions bot added enhancement New feature or request and removed enhancement New feature or request labels Dec 18, 2025

ogabrielluiz changed the base branch from main to feat/loop-subgraph December 18, 2025 12:54

github-actions bot added enhancement New feature or request and removed enhancement New feature or request labels Dec 18, 2025

ogabrielluiz added 2 commits December 18, 2025 10:53

feat(agent): add SharedContextComponent for multi-agent collaboration

8a799ce

feat(tests): add integration and unit tests for SharedContextComponent

41c26f6

github-actions bot added enhancement New feature or request and removed enhancement New feature or request labels Dec 18, 2025

ogabrielluiz added 5 commits December 26, 2025 09:51

feat(gitignore): add .claude to ignore list

51f157e

fix(frontend): reorder imports for consistency in SharedContextEvents…

d519e33

…Display component

feat(loop): update LoopComponent to enhance execution flow and add ne…

b2582f3

…w utility functions

ogabrielluiz changed the title ~~feat: Agent Blocks - composable agent building primitives~~ feat: Agent Blocks for composable agent Jan 7, 2026

github-actions bot added enhancement New feature or request and removed enhancement New feature or request labels Jan 7, 2026

This was referenced Jan 7, 2026

feat: add NotepadComponent #5845

Closed

feat: add new agent components for context management and action routing #5319

Closed

github-actions bot added enhancement New feature or request and removed enhancement New feature or request labels Jan 8, 2026

Merge feat/loop-subgraph into feat/agent-blocks

e69fdb1

- Fix extract_loop_output to handle dict outputs from model_dump() - Add loop subgraph execution infrastructure - Preserve component configuration across subgraph deepcopy

github-actions bot added enhancement New feature or request and removed enhancement New feature or request labels Jan 8, 2026

[autofix.ci] apply automated fixes

45ea79a

github-actions bot added enhancement New feature or request and removed enhancement New feature or request labels Jan 8, 2026

[autofix.ci] apply automated fixes (attempt 2/3)

a81922c

github-actions bot added enhancement New feature or request and removed enhancement New feature or request labels Jan 8, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat: Agent Blocks for composable agent #11055

feat: Agent Blocks for composable agent #11055

ogabrielluiz commented Dec 17, 2025 •

edited

Loading

Uh oh!

coderabbitai bot commented Dec 17, 2025

Review skipped

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

feat: Agent Blocks for composable agent #11055

Are you sure you want to change the base?

feat: Agent Blocks for composable agent #11055

Conversation

ogabrielluiz commented Dec 17, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Key Features

Architecture

Memory Integration

Benchmark Results (vs other agent frameworks)

Test plan

Uh oh!

coderabbitai bot commented Dec 17, 2025

Review skipped

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

ogabrielluiz commented Dec 17, 2025 •

edited

Loading