# Deep Agents vs. Plain ReAct: We Measured When the Harness Pays Off

2026-07-19 · Source: AI on Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Advanced, medium

Summary

A benchmark study compared LangChain's DeepAgents harness against a plain ReAct loop using DeepSeek V3.1 on 12 e-commerce tasks, graded by GPT-5.6 Luna. While DeepAgents passed 7/12 tasks to ReAct's 6/12, the deep arm incurred 3.6 times more tokens and roughly 6 times the wall-clock cost. The analysis revealed that plain ReAct either excelled on simple lookups (4/4 tasks, 1,776 avg tokens) or "detonated" on complex aggregations, often looping until step caps or silently quitting. DeepAgents, conversely, often fabricated confident but incorrect answers for tasks requiring complex joins, costing significantly more (e.g., 4.27M tokens for a failed monthly report). The study found that tool-to-task fit is paramount, with the harness offering little benefit if underlying tools are insufficient, and its "context offloading" features largely unused.

Key takeaway

For AI Engineers evaluating agent architectures, prioritize robust tool development over complex harnesses. If your tasks require complex data operations like database joins, invest in dedicated aggregation endpoints or code execution capabilities for your agent's tools. Deploying delegation with subagents demands verifiable subtasks; otherwise, you risk generating confident, fabricated answers. Always grade your agent's outputs against ground truth, as mere completion status does not guarantee correctness.

Key insights

Tool-to-task fit is more critical than agent architecture; inadequate tools lead to failure or confident fabrication.

Principles

• Plain ReAct excels when tasks fit tools.
• Harnesses amplify tool-to-task fit.
• Delegation requires verifiable subtasks.

Method

The study benchmarked DeepAgents against plain ReAct using DeepSeek V3.1 on 12 e-commerce tasks, graded by GPT-5.6 Luna against programmatic ground truth.

In practice

• Prioritize tool development over architecture.
• Grade agent outputs against ground truth.
• Use plain ReAct for simple, tool-fitting tasks.

Topics

• LLM Agents
• Agent Architectures
• LangChain DeepAgents
• ReAct
• Agent Benchmarking
• Tool Use

Best for: AI Architect, Research Scientist, AI Engineer, Machine Learning Engineer, AI Scientist

Related on AIssential

• The Anatomy of an Agent Harness — The agent harness, not the LLM, is the critical infrastructure enabling robust, multi-step AI agent performance.
• Tools vs. Subagents: Building Effective AI Agents Without Over-Engineering — Choose between tools for deterministic execution and subagents for complex, multi-step reasoning in AI agent architectures.
• DecisionBench: A Benchmark for Emergent Delegation in Long-Horizon Agentic Workflows — Emergent delegation in LLM agents is best evaluated via process-level metrics, as end-task quality alone obscures orchestration signals.
• EP218: The Typical AI Agent Stack, Explained — AI agents are complex systems built on a multi-layered architecture, not just an LLM and a prompt.
• not much happened today — AI development is converging on durable, inspectable agent execution, pretraining efficiency, and enterprise-grade security.
• Get to your first working agent in minutes: Announcing new features in Amazon Bedrock AgentCore — AgentCore simplifies AI agent development by abstracting infrastructure, enabling rapid prototyping and seamless production deployment.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI on Medium.