# This Week in SWE 2026.46: Agents must prove their work. Engineers build the safety rules.

First issue. This week, engineers make agents prove their work. They also remove dangerous commands from tools, write short human introductions on pull requests, and make skill files shorter. New tools: a local AWS emulator and Code Mode in Claude Code.

## 1. Write a short human introduction on agent pull requests 🧪

- **What:** Simon Willison writes a short introduction himself. He hides the long agent text in a block that opens when you click it.
- **Before:** You paste the full agent description into the pull request.
- **Now:** You write one or two sentences yourself. You put the agent text in a <details> block.

Words to know:
- **Agent:** An AI tool that does a coding task in many steps without help.
- **Pull request (PR):** A request to add your code change to the shared code. Other people examine it first.
- **<details> block:** An HTML element. It hides text until the reader clicks it.

*Method · experimental · code-review, pull-requests, ai-agents*

People: Simon Willison

Sources:
- [Hide your slop PR descriptions behind <details><summary>](https://simonwillison.net/2026/Nov/9/details-summary/), 2026-11-09 (primary)

Go deeper:
- Read: https://simonwillison.net/2026/Nov/9/details-summary/
- Try (15 min): On your next agent pull request, write the first two sentences yourself. Put the agent description in a <details> block.

## 2. Remove dangerous commands from the tool. Do not only tell the agent. 🧪

- **What:** Geoffrey Huntley gives every machine one changed version of git. This version cannot do a force push.
- **Before:** You tell the agent not to run dangerous commands.
- **Now:** You remove the command from the tool. You also add hooks, static checks and property-based tests.

Words to know:
- **Force push:** A git command that replaces the history on the server. It can delete the work of other people.
- **Nix overlay:** A Nix setting. It replaces a standard package with your changed version on each machine.
- **Git hook:** A script that git runs automatically before or after a command.
- **Property-based test:** A test that tries many random inputs. It checks a rule that must always be true.

*Method · experimental · agent-guardrails, nix, security, testing*

People: Geoffrey Huntley

Sources:
- [the answer to ai is engineering](https://ghuntley.com/answer/), 2026-11-11 (primary)
- [don't waste your back pressure](https://ghuntley.com/pressure/), 2026-01-17

Go deeper:
- Read: https://ghuntley.com/answer/
- Try (15 min): Find one dangerous command that your agent ran, for example a force push. Block it in the tool or in a git hook.

## 3. Ask the agent to prove that the change works 🧪

- **What:** Thorsten Ball (Amp) asks agents to prove that a change works from start to end. He says it changed how his team builds software.
- **Before:** You read the diff to find bugs.
- **Now:** The agent tests the change from start to end. It gives you screenshots, video or logs. Then you review.

Words to know:
- **Diff:** The list of lines that a change adds and removes.
- **End-to-end test:** A test of the full flow, as a real user does it. It does not test only one part.
- **Amp:** An AI coding agent product. Thorsten Ball works on it.

*Practice · experimental · verification, code-review, ai-agents*

People: Thorsten Ball

Sources:
- [Joy & Curiosity #105](https://registerspill.thorstenball.com/p/joy-and-curiosity-105), 2026-11-08 (primary)
- [Time Capsule 6: How Thorsten Ball Uses Amp](https://ampcode.com/time-capsules/6), 2026-10-13
- [AI Engineer: Your PR Says the Tests Passed. Prove It. (Stephen Chin & Cassandra Chin)](https://www.youtube.com/watch?v=gbeuC0ud1p4), 2026-11-10

Go deeper:
- Read: https://registerspill.thorstenball.com/p/joy-and-curiosity-105
- Watch: https://ampcode.com/time-capsules/6
- Try (15 min): At the end of your next agent task, tell the agent to test the change from start to end and to show proof.

## 4. Tell the agent who owns the code to control refactoring 🧪

- **What:** Anthropic guide for Claude Fable 5.2: one ownership sentence increased removal of duplicate code from below 10% of runs to about 75%.
- **Before:** You use scope rules such as 'no unrelated refactors'.
- **Now:** You add an owner sentence when you want a new structure. You omit it for small changes. You always check the diff.

Words to know:
- **Refactor:** To change the structure of code. The behaviour stays the same.
- **Prompt:** The instructions that you give to an AI model.
- **Duplicate code:** The same logic written in more than one location.

*Method · experimental · prompting, claude, refactoring*

People: Simon Willison

Sources:
- [Prompting Claude Fable 5.2](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-2), No date on page. Model launched 2026-11-09; earliest dated quote of this page is Simon Willison, 2026-11-10. (primary)
- [Quoting Prompting Claude Fable 5.2](https://simonwillison.net/2026/Nov/10/prompting-claude-fable-52/), 2026-11-10
- [Introducing Claude Fable 5.2 and Claude Mythos 5.2](https://www.anthropic.com/claude-fable-and-mythos-5-2), 2026-11-09

Go deeper:
- Read: https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-2#say-who-owns-the-code
- Debate: https://news.ycombinator.com/item?id=50429068
- Try (15 min): Do the same small cleanup task two times. Add an owner sentence only the second time. Compare the two diffs.

## 5. AWS Emulator: agents can test cloud code on a local machine 🧪

- **What:** AWS released a free emulator for more than 200 services. It works offline. Agents built it with tests first checked on real AWS.
- **Before:** You use mocks, or you test on a paid AWS account.
- **Now:** Tests run on a local emulator, so agents can check their own cloud code. Engineers own the specs, the tests and the merge rules.

Words to know:
- **Emulator:** Software that behaves like a different system. This one behaves like AWS on your machine.
- **Mock:** A simple false part that gives fixed answers in a test.
- **Merge rule (merge gate):** A check that must pass before code goes into the main branch.

*Tool update · experimental · aws, testing, ai-agents*

People: Swami Sivasubramanian, Werner Vogels, Matt Goldberg

Sources:
- [AWS Emulator and the engineer's job when agents write the code](https://www.allthingsdistributed.com/2026/11/aws-emulator-and-the-engineers-job-when-agents-write-the-code.html), 2026-11-11 (primary)
- [aws-emulator](https://simonwillison.net/2026/Nov/11/aws-emulator/), 2026-11-11

Go deeper:
- Read: https://www.allthingsdistributed.com/2026/11/aws-emulator-and-the-engineers-job-when-agents-write-the-code.html
- Debate: https://news.ycombinator.com/item?id=50456631
- Try (15 min): List the tests in one service that use AWS. Mark the tests that can run on a local emulator.

## 6. Claude Code Code Mode: one script runs many tool calls 🧪

- **What:** From version 2.1.323, Claude can write one JavaScript script that runs many tool calls, in sequence or at the same time. Experimental.
- **Before:** The model makes one tool call for each step and reads each result.
- **Now:** You set CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1. One script then does repeated edits and filtered searches. Terminal sessions only.

Words to know:
- **Tool call:** A request from the model to run a tool. Examples: read a file, run a command.
- **Environment variable:** A named setting that you give to a program when it starts.
- **Experimental:** Not final. The feature is off by default and can change.

*Tool update · experimental · claude-code, tool-calling, experimental*

Sources:
- [Claude Code CHANGELOG.md (2.1.323)](https://raw.githubusercontent.com/anthropics/claude-code/main/CHANGELOG.md), 2026-11-12 (primary)
- [Run tool calls from code with Code Mode](https://code.claude.com/docs/en/code-mode), No publish date on page. Page metadata says last modified 2026-11-13.

Go deeper:
- Read: https://code.claude.com/docs/en/code-mode
- Try (15 min): Start a terminal session with CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1. Ask for a change that repeats in many files. Press Ctrl+O to read the script.

## 7. Test each line of your SKILL.md. The model knows much of it. 🧪

- **What:** Nikolay Petrov (Intel) tests the model with and without each claim in a skill. His first skill was 70% known. Source: talk description.
- **Before:** Skill files only become longer, because you think it is a risk to delete a line.
- **Now:** You divide the skill into small claims and test each one. You keep only the lines that change the result.

Words to know:
- **Skill (SKILL.md):** A file of instructions that an agent loads for one type of task.
- **Claim:** One small statement in the skill that you can test.
- **Ablation:** A test where you remove one part and measure the change.

*Method · experimental · skills, context, evals*

People: Nikolay Petrov, Lydia Hallie

Sources:
- [AI Engineer: Your SKILL.md Is 70% Stuff the Model Already Knows (Nikolay Petrov, Intel)](https://www.youtube.com/watch?v=yzx4IuEh_tU), 2026-11-11 (primary)
- [AI Engineer: Rethinking Claude Code Best Practices (Lydia Hallie, Anthropic)](https://www.youtube.com/watch?v=vlfCjWuv4es), 2026-11-12

Go deeper:
- Watch: https://www.youtube.com/watch?v=yzx4IuEh_tU
- Try (15 min): Open your longest skill or CLAUDE.md. Delete five lines that the model possibly knows. Do one real task again and compare the result.