<aside> 💡

This blog will present our new multi-agent library, ContextAgent. It is a minimal, context-central multi-agent framework with PyTorch-like API. We want to build intelligent agent systems with minimal code.

</aside>

A context-centric multi-agent framework that integrates seamlessly with OpenAI Agents SDK. PyTorch-style definitions, open-box tracing, configuration, and rich printer utilities. Easy to customize; batteries included.

• Context-first: agent = context + LLM. Agents are instantiations of context bound to an LLM.
• Composable: profiles, pipelines, agents, and tools are loosely coupled.
• Open tracing: spans, groups, and status panels for every step.
• Providers: OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter, Bedrock via LiteLLM.

ContextAgent began as a response to a pattern we kept seeing in multi-agent projects: teams were creating dozens of bespoke agent types, stitching them together with brittle orchestration code, and manually juggling prompt state. Despite the tooling looking sophisticated, the real differentiator was always the context each agent carried. We decided to build a framework that embraces that reality. ContextAgent is a context-centric agent framework that puts context engineering front and center, so you can build reliable, flexible AI systems with far less ceremony.

Why a Context-First Architecture Matters

Modern agents succeed or fail based on the context they see. When organizations prototype “multi-agent” systems, they often realize that their so-called agents are simply the same foundation model wearing different prompt templates, state, and run-time data. Treating each one as a unique entity encourages boilerplate, hides the shared mental model, and makes experimentation slow.

We believe the correct abstraction is:

• The agent is a thin execution shell: an LLM client and a context compiler.
• The context is the dynamic payload that encodes task shape, role, tools, and memory.
• The pipeline is the choreography that decides how contexts evolve over time.

By designing for context first, we get three compounding benefits:

1. Faster iteration. Prompt changes, tool outputs, and runtime knowledge can be injected or replaced without redefining class hierarchies.
2. Safer reasoning. Prompts and state live in one place, making it easier to review, enforce policies, and trace unexpected behavior.
3. Better reuse. Contexts are composable assets. You can share them across teams, load them from configs, or compile them from templates and data sources on demand.

Traditionally, multi-agent frameworks over-index on agent types and under-index on context lifecycle. ContextAgent flips that emphasis. We intentionally decouple agent and context so you can reason explicitly about how information flows, why a certain LLM call behaves the way it does, and what levers you can pull to change outcomes.