Blog

Working with AI coding agents

Practical notes for teams that build with Claude Code, Codex and Cursor: reviews, rule files, hooks, MCP and security. Memmo is built by agents that use Memmo, so some posts are the lessons behind it.

• Your coding agent's permission rules had gaps. Update this week

  This week's Claude Code releases fixed several cases where permission and deny rules didn't apply, and a Codex change makes CI runs ask for explicit project trust. What changed, why permission rules aren't a security boundary, and how to keep agents patched.

  News · 5 min read

• Coding agents often fix the bug early, then keep going

  A new paper finds coding agents often reach a correct fix on their first edit, then spend a third to over half of the run on extra steps, and sometimes undo the fix. What it means for cost, and how to give agents a clear finish line.

  News · 5 min read

• Your team writes AI code faster than it can check it

  A new Sonar survey of 1,026 developers says AI now writes or heavily assists 65% of committed code, 76% generate it faster than they can verify it, and only 23% have rules for it. What that means for review, and what to put in place.

  News · 5 min read

• Two new top coding models in a week. Should your team switch?

  Anthropic released Claude Fable 5.2 and Google opened up Gemini 4 Carbon within a day of each other. Both vendors still say their cheaper model should do most of your work. How to decide, and how to test a new model on your own code.

  News · 5 min read

• Coding agents gave away a secret when other agents asked

  In a sandboxed experiment, coding agents told to guard a token shared it with a forum of other agents and ran their install script, and most didn't mention it afterwards. What the test found, its limits, and what it means for agents you let roam.

  News · 5 min read

• Your agent's "don't ask again" list may hold your passwords

  Researchers found over 1,100 public copies of Claude Code's local settings file containing secrets, saved there when developers approved a command for good. How it happens, how to check your repos, and what to change.

  News · 5 min read

• Apple cut its engineers' AI budgets. Does your team have one?

  Apple reportedly lowered some engineers' daily AI spending cap from $300 to $200, and it is not the only company setting limits. Agent spend is becoming a line item. How to budget it without slowing your team down.

  News · 5 min read

• Your agent's context can leak through its requests

  A Claude Code user found their email address in the User-Agent header of the agent's web requests. The fix is coming, but the lesson applies to every coding agent: anything in its context can end up in a tool call. What to keep out, and what to watch.

  News · 5 min read

• What 400 developers say about their AI agent setup

  A Hacker News thread asked developers to describe their AI agentic setup, and hundreds answered. Most mix tools, a quarter run agents in parallel, and the best tips are surprisingly plain. What a team lead can take from it.

  News · 5 min read

• Postman installed itself into Claude Code and Codex. Who else can?

  An update to Postman's desktop app added plugins, a startup hook and an MCP server to developers' coding agents without asking. What happened, how to check your machines, and why agent config now needs the same care as your shell profile.

  News · 5 min read

• Your coding agent may be installing MCP servers nobody maintains

  New research shows coding agents setting up archived MCP reference servers from memory, including one with a remote code execution bug that will never be fixed. How to find them on your team's machines, and what to standardise on instead.

  News · 5 min read

• OpenAI rebuilt Git for its agents: what your team should take from it

  Codex pushed so many branches, fetches and merges at OpenAI that merges sometimes failed more than one time in ten. Its answer was a new Git backend called Pando. You won't need one, but your repo, CI and reviewers will feel the same pressure.

  News · 5 min read

• The backdoored Claude Code extension: what it means for your team

  For eleven hours on 29 October, Cursor's extension marketplace served a tampered Claude Code extension that planted an SSH key on developer machines. What happened, how to check your team, and why extensions now belong in your supply chain policy.

  News · 5 min read

• How to review a pull request written by an AI agent

  Agent pull requests look finished, so polish no longer tells you anything. Here is what to read first, the mistakes agents make most often, and a checklist you can use on the next one.

  Code review · 6 min read

• CLAUDE.md, AGENTS.md, Cursor rules: one set of rules for every agent

  Claude Code, Codex, Cursor and Copilot each read a different file. Which file each agent reads, how to keep one source of truth, and how to stop the copies drifting apart.

  Agent setup · 5 min read

• How to write a task an AI coding agent can finish

  Most failed agent runs start with a vague task, not a weak model. A five-part template, a before and after example, and how to size work so the result is easy to review.

  Guide · 5 min read

• Claude Code hooks: guardrails that don't depend on the prompt

  An instruction is a request. A hook is a guarantee. Four hooks worth adding to a team repo: protect paths, format after edits, brief at session start, and don't stop on red tests.

  Agent setup · 6 min read

• MCP explained for engineering teams

  The Model Context Protocol lets coding agents use your tools and data. What it is, how local and remote servers differ, what teams connect, and how to choose servers without opening a hole.

  MCP · 5 min read

• Prompt injection and coding agents: a threat model for your team

  Coding agents read text written by strangers and can run commands. Where untrusted text gets in, what an attack looks like, and the layered defenses that don't rely on the model noticing.

  Security · 6 min read

• Running several coding agents in parallel without chaos

  One agent at a time leaves you waiting. Several in one checkout step on each other. Git worktrees, separate ports and databases, and choosing tasks that don't collide.

  Guide · 5 min read

• Tests first: getting code you can trust from an AI agent

  Agents are very good at making tests pass, sometimes by changing the tests. A tests-first workflow, the warning signs to look for, and how to protect the suite from the agent.

  Guide · 5 min read

• Context engineering for coding agents: what goes in the window

  Everything an agent knows about your task has to fit in its context window, and more is not better. Three tiers of context, and how to keep the window clean on long tasks.

  Guide · 5 min read

• How to measure whether AI coding agents help your team

  Lines of code and acceptance rates say little, and people's own sense of speed can be wrong. The delivery and quality signals worth tracking, and how to set a baseline first.

  Leadership · 5 min read

• Code review that knows your codebase

  AI reviewers read the diff. They don't know what your team decided, what broke last time, or what you already dismissed. Here is how Memmo grounds every pull request review in the team's approved memory.

  Memmo · 6 min read

• How we stop Claude Code repeating mistakes

  Every Claude Code session starts from zero, so the same mistakes come back. These are the four moments where we hand the agent what our team already learned, and the real bugs that stopped coming back.

  Memmo · 7 min read