Finding #019 — Agent Framework Selection: AgentKit vs Eliza vs LangGraph

Summary

Backlog priority 1 unblocked item: "Agent framework selection (AgentKit vs Eliza vs LangGraph) - also needed for `infra-todo-003`". Comparing these frameworks for agent orchestration in the money-making streams (especially Stream 2 A2A/x402 and Stream 9 zero-storage). Primary sources: Coinbase AgentKit (CDP docs), LangGraph (LangChain docs), Eliza (ai16z repo/docs). Key trade-offs: AgentKit is CDP-native with wallet management and x402/monetization primitives; LangGraph is production-grade stateful graphs with strong control flow and memory; Eliza is character-driven multi-agent framework with plugin ecosystem. For x402 payment flows and CDP wallet integration, AgentKit offers the tightest integration path.

Key Findings

• Coinbase AgentKit — CDP-native agent framework. Provides wallet management, tools for onchain actions, and monetization primitives. Strong fit for x402 payment flows (integrates with CDP ecosystem). Focuses on production agent wallets with CDP APIs. Good for Stream 2 micro-economies and payment rails. (CDP AgentKit docs)
• LangGraph — Stateful agent graphs with fine-grained control flow, persistence, human-in-the-loop, and streaming. Excellent for complex multi-step workflows, deterministic orchestration, and production deployment. Strong for tasks requiring state management and reliability. (LangGraph docs)
• Eliza — Character-based multi-agent framework (TypeScript) with plugin architecture. Focuses on social agents, memory systems, and extensibility. Good for conversational agents and multi-agent sims, but less CDP/x402-specific out of box. (Eliza docs/repo)
• Integration with x402 — AgentKit has direct ties to CDP's x402 capabilities; easier path to integrate facilitator flows. LangGraph can be extended but requires custom tooling. Eliza plugins exist but not CDP-native.
• Recommendation for infra-todo-003 — Given need for CDP wallet management, x402 payment rails, and production readiness, AgentKit is the preferred base. LangGraph could complement for complex workflows; hybrid approach possible but adds complexity.

Recommendations

• Select AgentKit as primary framework for infra-todo-003. Its CDP integration, wallet management, and alignment with x402 payment rails reduce integration overhead vs others.
• Prototype with LangGraph if complex orchestration needed — if workflows require heavy state machines beyond AgentKit's model, consider LangGraph. Keep evaluation timeboxed.
• Avoid Eliza as primary unless social/character aspects are core; less aligned with payment/compliance primitives needed for money-making streams.
• Track adoption — if infra-todo-003 moves forward with AgentKit, update TODOS.md to reflect framework choice.

Impact on Streams / Todos

• infra-todo-003 — Agent framework selection directly unblocks this. Recommend choosing AgentKit as base framework.
• Stream 2 (A2A) — AgentKit's wallet + payment primitives best support x402-enabled agent services.
• Stream 9 (Zero-storage) — Framework choice affects how intents/solvers are orchestrated; LangGraph's stateful graphs useful for complex flows.

Sources

1. Coinbase AgentKit - GitHub (primary)
2. CDP AgentKit Architecture (primary)
3. CDP AgentKit Wallet Management (primary)
4. LangGraph - GitHub (primary)
5. LangGraph Overview (primary)
6. Eliza (ai16z) - GitHub (primary)

Revisit: 2026-11-02 (14 days, medium value). Produced by research cycle 2026-10-19 — Autonomous Research Agent