2025 was the year of agents, what's coming in 2026?

January 9, 2026

51 min episode · 2 min read

Episode

51 min

Read time

2 min

AI-Generated Summary

Published Jan 11, 2026

Key Takeaways

✓Agent Implementation Success: Effective AI agents require domain expertise to configure prompts, select data sources, and integrate tools like MCP servers. Organizations lacking this expertise face high failure rates, with Gartner predicting 40% of projects will fail by 2027 despite 11% having agents in production.
✓Reasoning Model Trade-offs: Models like Claude Opus 4.5 and OpenAI o1 generate intermediate reasoning tokens before final outputs, enabling senior-level coding capabilities. However, each reasoning token requires separate model inference runs, dramatically increasing latency and computational costs for production applications.
✓Power Infrastructure Bottleneck: GPU availability no longer limits AI advancement; power consumption does. Speculators purchase decommissioned power plants anticipating reactivation needs. Energy requirements now drive geopolitical policy decisions, with AI infrastructure investments facing community resistance over power demands and environmental impact.
✓AI Engineering Skill Set: The emerging valuable role combines data science, software development, and system architecture to build MCP servers, connect databases, integrate RAG systems, and orchestrate multiple AI services. This integration expertise remains complex enough to resist automation for years.

What It Covers

Hosts Daniel Whitnack and Chris Benson review 2025 as the year AI agents emerged, examining successful implementations, reasoning model advances, infrastructure challenges, and predictions for 2026's increasingly complex AI ecosystem.

Key Questions Answered

•Agent Implementation Success: Effective AI agents require domain expertise to configure prompts, select data sources, and integrate tools like MCP servers. Organizations lacking this expertise face high failure rates, with Gartner predicting 40% of projects will fail by 2027 despite 11% having agents in production.
•Reasoning Model Trade-offs: Models like Claude Opus 4.5 and OpenAI o1 generate intermediate reasoning tokens before final outputs, enabling senior-level coding capabilities. However, each reasoning token requires separate model inference runs, dramatically increasing latency and computational costs for production applications.
•Power Infrastructure Bottleneck: GPU availability no longer limits AI advancement; power consumption does. Speculators purchase decommissioned power plants anticipating reactivation needs. Energy requirements now drive geopolitical policy decisions, with AI infrastructure investments facing community resistance over power demands and environmental impact.
•AI Engineering Skill Set: The emerging valuable role combines data science, software development, and system architecture to build MCP servers, connect databases, integrate RAG systems, and orchestrate multiple AI services. This integration expertise remains complex enough to resist automation for years.

Notable Moment

One host describes spending weeks researching an autonomy project, then crafting a detailed prompt that generated six weeks worth of production-quality code in six minutes, representing a transformative workflow shift enabled by late 2025 model capabilities.

Know someone who'd find this useful?