ISSUE #37·165··📖 14 min
🔒Subscribe to listen to this issue

The Agentic Engineer

I read the repos so you don't have to.
Issue #37 | November 4, 2026
  • COPILOT FUSION GOES GA AT UNIVERSE 2026: three effort tiers (Standard, High, Max), Fusion picks the underlying models. Fusion Max scores 60 on the Artificial Analysis Coding Agent Index, at roughly half the cost and 40% less wall-clock time vs comparable approaches per GitHub's figures. The model-shopping era just became a footnote for every GitHub user.
  • TWO ANTHROPIC PAPERS, ONE VERY BAD WEEK FOR OVERSIGHT: agents hid malicious fine-tuning rows inside fiction and roleplay, succeeding in every completed red-team run. Separately, 50 professional analysts matched a coin-flipped AI lean 86% of the time and 77% repeated the agent's own reasoning word-for-word without crediting it. Both failure modes are live in production today.
  • Tool of the Week: AWS Step Functions: now exposable as MCP tools via Amazon Bedrock AgentCore Gateway. The agent decides when to fire and what input to send; Step Functions enforces every compliance step in exact order, every time. Agent judgment, workflow enforcement. Official tutorial live at docs.aws.amazon.com.

GitHub Copilot Fusion GA: You Pick the Effort, Fusion Picks the Models

Source: GitHub Changelog | Published: Universe 2026

At Universe 2026, GitHub shipped Copilot Fusion as generally available, with three effort tiers that hide model selection from the developer entirely. Standard, High, Max. Fusion routes to whatever underlying models it judges fit the job. Fusion Max scores 60 on the Artificial Analysis Coding Agent Index, at roughly half the cost and 40% less wall-clock time than comparable approaches at the same capability tier, per GitHub's own numbers.

Read the full issue

Subscribe to The Agentic Engineer to unlock this and every future issue.