Context Switching Costs Developers 40% of Their Productivity

▶ Watch (02:01)

Building software means touching an IDE, a GitHub repo, a generative AI chat window, a deployment tool, and a monitoring dashboard, all to ship one feature. Each switch costs time. Sius put a number on it:

“It’s been stated that you lose up to 40%” of your productivity just by switching contexts too much. — Jemiah Sius

That 40% is not a rounding error. It is the drag built into every standard SDLC. Reducing it means reducing the switches, which means agents that cross tool boundaries on the developer’s behalf.

From Generating to Acting: What Agentic AI Actually Does

▶ Watch (05:13)

ChatGPT and GitHub Copilot generate content, write code, and answer questions. They stop there. Once a workflow needs something to actually happen, generative AI hits a wall. Agentic AI moves past that wall by running multi-step reasoning on its own: gathering data, forming a plan, calling other tools or agents, executing tasks, and updating its approach based on feedback. Sius used Waymo as the reference: give it a destination, and it makes hundreds of routing decisions without asking again.

How Agents Find Each Other and Agree on a Protocol

▶ Watch (08:21)

Before two agents can cooperate, they go through three layers. First, discovery: one agent scans a service registry, reads agent cards (Google’s A2A protocol defines this format), and finds the right peer. Second, protocol negotiation: they agree on how to talk. MCP from Anthropic handles agent-to-tool calls. A2A from Google handles direct agent-to-agent messages without a middle framework.

“MCP right. We’ve heard that we’ve used some of that.” — Jemiah Sius

Third, transport: messages move as JSON or JSON-RPC 2.0 over HTTP or WebSockets, with error-handling instructions baked into each payload.

Two Agents Closing a Production Incident Without a Human Pager

▶ Watch (12:17)

A developer merges a PR. A GitHub Action fires a change event to New Relic AI, including commit SHA, service metadata, and author. New Relic AI traces errors, latency spikes, and newly exposed vulnerabilities down to the transaction level, then opens a detailed GitHub issue with stack traces, slow query analysis, and vulnerability findings. The GitHub Copilot + New Relic agent closed-loop demo (12:17) shows what happens next: Copilot Coding agent reads the issue, writes code and tests, opens a PR.

“You keep a human in the loop.” — Jemiah Sius

The developer approves. The cycle restarts.

Notable Quotes

It’s been stated that you lose up to 40% Jemiah Sius · ▶ 3:40

MCP right. We’ve heard that we’ve used some of that. Jemiah Sius · ▶ 9:30

You keep a human in the loop. Jemiah Sius · ▶ 15:22

Key Takeaways

  • Constant tool-switching costs developers up to 40% of their productivity across the SDLC.
  • Agentic AI differs from generative AI by taking autonomous action, not just producing content.
  • MCP handles agent-to-tool calls; A2A handles direct agent-to-agent coordination without a framework.