The PracticeThe PracticeThe PracticeOperating infrastructure

The Practice · The Lookout

How to think about agents and harnesses


What the frontier labs say their own models and agents can do, read for what it changes in a working practice.

By Jake Griffith

The Lookout · I of V · 6 entries

Claude Opus 5.5 System Card

Anthropic, September 22, 2026

What it says

The frontier model's own account of its gains in coding, agentic, and computer-use work, with the safeguards it ships under.

What it changes

The capability a client's install actually runs on moved with this release; read the agentic and computer-use sections before promising what a workflow can do unattended.

Claude Sonnet 5.5 System Card

Anthropic, September 28, 2026

What it says

The working-tier model now carries agentic capability and measured resistance to prompt injection in coding and browser work.

What it changes

The model most installs run on day to day; the injection-resistance section is the one to cite when a team asks whether an agent can be trusted in a browser.

GPT-6 Sol and Luna

OpenAI, September 2026

What it says

Two tiers at lower price points beside GPT-6 Astra, with cached-input discounts and the lab's own figures on reduced deception in coding tasks.

What it changes

The cost line in every ChatGPT-side install; re-price the per-seat arithmetic in the Fit Finder's assumptions.

Dots, OpenAI's personal agents

OpenAI, September 29, 2026

What it says

Always-on agents in ChatGPT that take on ongoing work on a user's behalf across the apps you connect, each with its own cloud computer, powered by GPT-6 Astra, on paid Pro and Business Premium plans, with specialist dots previewed for defined roles inside an organization.

What it changes

A vendor has shipped the personal-agent-that-acts pattern, so the "what did the agent actually do" question now has a precedent; the audit-trail rule in every install is no longer a house preference.

Building Effective Agents

Anthropic, December 2024

What it says

Most production "agents" are simple composable workflows, and the simplest architecture that works beats the elaborate one.

What it changes

The catalog's design rule in one essay; read it before buying any orchestration product.

Claude documentation: prompting and evaluation guides

Anthropic, ongoing

What it says

Concrete rubrics, reason-then-verdict grading, and prompt structure, from the publisher.

What it changes

The Grading Room's method is drawn from here; every evaluator rubric should be checked against it.

Next: The papers behind the instruments

Educational, not professional advice. Nothing on this page is affiliated with or endorsed by its publisher.