I Kept Comparing Models. I Was Ignoring Half the System.

What coding agents, MCP, and context systems taught me about harness engineering

Sep 2 • Mayank Bohra

2

1

July 2026

June 2026

I Built a Cross-Session Memory Layer for Claude. Here Are the Four Ways It Broke.

Drew Breunig named four ways long context fails: poisoning, distraction, confusion, clash. I hit all four shipping Context Hub and Multicast. The bugs…

Jun 15 • Mayank Bohra

2

The NSA Wrote an MCP Security Playbook. I Run an MCP Gateway. The Real Risk Is the Boring One.

Everyone's hunting exotic MCP exploits. The hole that actually burns you is one server holding every credential, wired up with zero auth. Here's what…

Jun 14 • Mayank Bohra

3

1

Your AI Agent Doesn't Have a Memory Problem. It Has a Context Rot Problem.

I've run a cross-session memory layer for Claude for months. The fix was never a bigger window. It was learning to throw context away on purpose.

Jun 11 • Mayank Bohra

2

2

When Your AI Setup Learns and You Do Not

After a year of daily AI tooling I could search every idea I had saved and explain almost none of them without the tool open. Here is the one test I now…

Jun 4 • Mayank Bohra

2

1

April 2026

January 2026