---
title: 'In the News: November 19, 2026 (Evening)'
description: "HubSpot explains Aviator, the shared Java harness behind its agents. Claude Code now lets an erroring hook's call through in Code Mode."
canonical_url: 'https://darkfactory.dev/news/2026-11-19-evening'
markdown_url: 'https://darkfactory.dev/news/2026-11-19-evening.md'
collection: news
date_published: '2026-11-19T19:05:00-05:00'
date_modified: '2026-11-19T19:05:00-05:00'
---

# In the News: November 19, 2026 (Evening)


HubSpot has published how it built one agent harness for hundreds of teams, and what that harness does when a single assistant needs more than 3,000 tools. Separately, today's Claude Code release changes what happens when a guard hook crashes inside Code Mode, which matters to anyone using hooks as a safety check.

## 1. HubSpot runs its agents on one shared Java harness

**[Aviator: HubSpot's Agent Harness](https://hubspot.engineering/blog/aviator-hubspots-agent-harness/)** · Ze'ev Klapow, HubSpot · HubSpot Engineering, November 19, 2026

HubSpot started looking for a common agent framework in June 2025 because each team was making its own choices and repeating the same mistakes with prompt caching, tool calling and context management. "Writing the first version of an agentic loop is easy," Klapow writes. "Scaling a harness to hundreds of teams is a different problem." Most of HubSpot's backend is Java, so the company built Aviator as a Java library with a target of an agent in under 20 lines. Agents call existing internal RPCs as tools.

Aviator stores every conversation in its own provider-neutral transcript format, so an agent can switch models mid-conversation (tool-call IDs are remapped and incompatible reasoning blocks removed), hand a transcript to another agent, or save it as JSON and resume days later on another machine. An agent can pause while it waits for a human to approve a tool call. Tools can be marked deferred and found through search: the post says Sidekick, the internal engineering assistant, "needed access to over 3,000 tools." Skills extend the SKILL.md format with `load-tools` and `suggested-tools` fields so a skill brings its tools with it. Code runs in Python and Bash sandboxes compiled to WASM that ship inside the Java library, reach only egress points HubSpot defines, and start fast enough to run "tens of thousands of short-lived environments."

The post reports that Sidekick Code, HubSpot's cloud coding agent on Aviator, "now merges more than 2,500 pull requests a week." It shows an adoption chart with the counts left out. The figures are HubSpot's own, and the post includes no code or data a reader could use to check them.

**Why it matters:** This is a large company's case for owning the harness layer itself rather than letting each team pick one. If you run agents on more than one model, the parts to copy are the neutral transcript, which makes model switching and pause-and-resume ordinary, and loading tools on demand once the list grows past what fits in a prompt.

## 2. Claude Code: a crashing hook in Code Mode no longer blocks the call

**[Claude Code v2.1.329](https://github.com/anthropics/claude-code/releases/tag/v2.1.329)** · Anthropic · GitHub release notes, November 19, 2026

The release changes Code Mode "so a PreToolUse hook that errors no longer refuses a script's tool call, as outside Code Mode." Hooks that must stop the call when they fail now need `"onFailure": "block"` on the command, HTTP or MCP tool hook. The trust dialog now warns where a folder's hooks are declared, and Claude Desktop and other non-interactive hosts show a warning when hooks in managed settings fail to load.

The release also removes the automatic retry at the limit named in a `max_tokens` error that version 2.1.326 had added. Behind a gateway whose model takes fewer output tokens, the notes say to set `CLAUDE_CODE_MAX_OUTPUT_TOKENS` to that limit. And it fixes the Agent tool's `effort` parameter being ignored for named teammates in agent teams.

**Why it matters:** If you use a PreToolUse hook as a guard in Code Mode and it can crash, a crash now lets the call through unless you set `"onFailure": "block"`. Check which of your hooks are meant to fail closed before you upgrade.
