The Coolest Agents I've Built So Far

What It Covers

Host NLW runs 16 of his 2025 AI builds through a March Madness-style bracket tournament, covering agents built with OpenClaw, Claude Code, and Perplexity. Mycroft, a Slack-based digital Chief AI Officer that builds continuous company-wide AI strategy roadmaps, wins the tournament over Chucky, an interactive agent portfolio representative.

Key Questions Answered

• •Agentic shift timeline: The transition from standard AI tools to agent-based workflows accelerated dramatically in the three to four months preceding March 2025, driven by OpenClaw, Claude Code, Codex, and Perplexity Computer. Builders should audit current workflows now and identify which repetitive research, project management, or discovery tasks can be handed to a persistent autonomous agent.
• •Persistent AI strategy over one-time consulting: Rather than deploying AI assessments as periodic engagements, the Mycroft model runs continuously in Slack, updating company-wide AI roadmaps across six vectors: use cases, systems, data integration, outcomes, people, and governance. Organizations can replicate this by assigning a dedicated agent to maintain a living strategy document rather than a static quarterly report.
• •Agent portfolio representation via Chucky: When demonstrating AI-building skills to clients or employers, static resumes and portfolios fall short. The Chucky model deploys an interactive agent that fields questions, surfaces screenshots, links to live tools, and visualizes the full ecosystem of builds. Builders should consider creating a conversational representative rather than a PDF portfolio for client outreach.
• •24/7 autonomous research as highest-utility OpenClaw use case: The Widi Radars researcher agent runs continuously, scanning studies, surveys, and reports to populate a tiered use-case database categorized as Primetime, Emerging, or Frontier. Teams tracking fast-moving domains like AI adoption should deploy a persistent research agent feeding a structured database rather than relying on manual literature reviews.
• •Power users average 3.5 models simultaneously: Monthly pulse survey data from the AIDB community shows the most active AI users employ an average of 3.5 different models, each selected for specific use cases. Practitioners should map their recurring task types, then deliberately assign the most capable model per task category rather than defaulting to a single general-purpose model for everything.