Sunday, July 5, 2026
The agent is the operator.
July 5 · 4 videos
Shipping is just the start.
Wandero AI uses agents to fix agents.
RELAI hit 97 percent success with VCL.
MCP is the new browser.
Clarity is now a survival skill.
“The failure is in the conversation, not the stack trace.”
The Missing Layer After Launch - Raphael Kalandadze, Wandero AI
Raphael Kalandadze · AI Engineer · 19 min
Watch on YouTube →Raphael Kalandadze explains why shipping agents is not the finish line. He details a meta-harness of specialized agents that monitor and repair production systems.
- Agent failures are often hidden in conversations rather than stack traces.
- Wandero AI uses a Log Monitoring Agent to draft PRs every 15 to 60 minutes.
- A Review Agent evaluates these PRs to prevent hallucinated fixes.
- The system produces a 10 times ratio of agent-generated PRs to human-generated ones.
- Humans move from manual debugging to high-level approval at the merge boundary.
- Operating an agent is itself an agent problem requiring reasoning over messy logs.
- A tight feedback loop for production health is a greater moat than the initial model.
Continual Learning for AI Agents: From Failures to Durable Improvements - Soheil Feizi, RELAI
Soheil Feizi · AI Engineer · 22 min
Watch on YouTube →Dr. Soheil Feizi introduces Verifiable Continual Learning to solve agent failures in production. This framework turns session logs into replayable simulations for durable improvement.
- Static benchmarks fail to capture the complexity of real-world agent interactions.
- VCL uses mock tools and synthetic personas to create replayable learning environments.
- A support agent improved from 78 percent to 97 percent success in one cycle.
- Optimizing the harness and memory is often faster than fine-tuning model weights.
- Verifiable updates prevent the regression common in manual prompt engineering.
- Verification acts as the bridge between development benchmarks and production reliability.
- Agents must move from static training to lifelong learning loops to remain relevant.
MCP Apps: Primitives, discovery, and the Future of Software - Pietro Zullo, Manufact, Inc
Pietro Zullo · AI Engineer · 28 min
Watch on YouTube →Pietro Zullo argues that MCP Apps are the new architecture for the web. He describes a shift where AI hosts like Claude and Cursor replace traditional dashboards.
- MCP Apps add a stateful, bidirectional UI layer to standard model connectors.
- The mcp-use framework has surpassed 8 million downloads across PyPI and npm.
- Users are migrating to AI-mediated workflows instead of navigating external websites.
- Claude now supports dynamic discovery to find the right MCP connector automatically.
- Privacy-preserving redaction allows users to see data that remains hidden from LLM providers.
- Market positioning is shifting as users buy products based on MCP server availability.
- The traditional dashboard-centric SaaS model is being bypassed by agentic hosts.
Your AI Product Will Fail Unless You Can Explain It - Veronica Hylak, Hey AI
Veronica Hylak · AI Engineer · 6 min
Watch on YouTube →Veronica Hylak addresses the communication gap facing technical AI founders. She provides a framework to make complex agentic products understandable to buyers.
- The 17-year-old test determines if a value proposition is too complex.
- Founders must master a 20-second pitch focusing on the wound and the fix.
- Great technology dies quietly if the market cannot understand its purpose.
- Analogies like smoke alarms for AI create instant contextual relevance.
- Transformation demonstrations should show concrete before-and-after scenarios in seconds.
- Clarity is now a prerequisite for funding and sales in a saturated market.
- Storytelling is a survival skill in the AI era rather than a secondary marketing task.
References
PeopleRaphael Kalandadze (@RaphaelKalan) · Harrison Chase · Soheil Feizi (https://x.com/FeiziSoheil) · Pietro Zullo (https://x.com/pietrozullo) · Idel Solomon · Paul Graham · Veronica Hylak
ToolsWandero AI · RELAI · MCP · mcp-use · ChatGPT · Claude · Cursor