The Agentic Engineer
I read the repos so you don't have to.
Issue #20 | July 8, 2026
TL;DR
- Better models are breaking third-party coding tools. Armin Ronacher documented how Opus 4.8 and Sonnet 5 invent fake schema fields when calling Pi's edit tool. RL training on Claude Code's built-in tools is making models worse at everything else. Every coding harness builder needs to pay attention.
- Last week you voted for agent sandboxing. OpenAI's codex-plugin-cc (25K stars) now lets you use Codex from inside Claude Code. The walls between competing agents are dissolving. Meanwhile Strix hit #1 trending with autonomous AI pentesting agents.
- AWS CloudFormation Express Mode is this week's Tool of the Week: 4x faster infrastructure deployments designed for the agent-driven development loop. Sub-minute feedback for iterating on infra. Free everywhere.
The Big One
Better Models, Worse Tools: SOTA Claude Models Break Third-Party Coding Harnesses
Armin Ronacher (creator of Pi) published a detailed investigation on July 4 that every coding harness builder should read. The short version: Opus 4.8 and Sonnet 5 invent extra fields when calling Pi's edit tool. The edits themselves are correct. But the models add garbage keys like requireUnique, oldText2, type, and kind to the tool call arguments.
Pi rejects the malformed calls. The model retries. Tokens burn. Users get frustrated.
Read the full issue
Subscribe to The Agentic Engineer to unlock this and every future issue.