How do teams protect quality and safety with AI tools?

GitLab’s AI-Assisted Development Playbook, added on 27 March 2026, splits agent use into five levels: Baseline, Pair, Conductor, Orchestrator and Harness. A repo should not move past Baseline until it reaches maturity level 2 on all four dimensions: CI and constraints, context and docs, testing depth, and review practice.

Each quarter Atlassian takes product managers and designers out of day-to-day work for four days to build with AI. Day one is inspiration (customer and internal showcases, outside speakers), day two is skills (reusable agents, evaluations, a dev environment), day three is building in one of seven groups with live mentoring, and day four is a demo day that includes what failed. In the latest week 77% of Product Management and 80% of Design attended.

AI and automation remove repetition that does not require human judgment.

A coding task is first attempted with a prompt in Warp.

In thoughtbot’s two-week TellaDraft client engagement, the team used small AI tasks, a CLAUDE.md repository context file, meaningful tests for every feature, and human review before each commit; this is not presented as a company-wide policy.

dxw’s guidance updated 23 September 2026 names company-administered tools; client data requires permission before use in a tool not provided by that client. Making decisions with AI is not allowed; external content and agentic actions need special care.

Bending Spoons’ 13 August 2026 SEC disclosure says Alt-Spooner operates with the access of the person it assists and their connected accounts. Introduced in Q2 and rolled out to all staff in early July, it reportedly processed over 100 billion tokens in its first three weeks.

Fleet expects engineers to use Claude Code in their daily workflow. In return, no AI-generated work (code, blog posts, handbook edits, customer emails) ships until both author and reviewer have read and understood it line by line; multiline product changes are run locally and the PR is opened from the person’s own GitHub account, not a bot.

Anyone contributing to Firefox with AI tools keeps patches small, must be able to understand and explain every line, and completes their own review before asking for one. Confidential or security-sensitive information is not given to external AI tools. The policy was published in November 2025; the self-review requirement was added in May 2026.

Niteo’s handbook Career section now includes “Use of AI Tools”: regardless of position or track, everyone is expected to be comfortable using LLMs and other AI tools — not to become an expert prompt engineer, but to stay curious about new tools, reduce repetitive work and keep improving their workflow. This sits inside the formal promotion/leveling-criteria document.

Engineers scaffold a new skill (a Claude-Code-style agent capability) under `products/{product}/skills/` with `hogli init:skill`, validate it with `hogli lint:skills`, and build it locally with `hogli build:skills`. On merge to master, CI automatically builds `dist/skills.zip` and publishes it to the `PostHog/skills` and `PostHog/ai-plugin` repositories, which PostHog Desktop and PostHog AI consume.

Netflix built an open-source agent (`oci-agent`) that automates the laborious steps of observational causal inference — covariate balance checking, propensity-score trimming, sensitivity analysis. The agent must produce auditable artifacts (plans, specs, plots, notebooks) that a human analyst can review and re-run; the final causal claim goes through human sign-off before it feeds a decision. Open-sourced in June 2026.

When Red Hat engineering found that its first AI policy, though sound on risk and legal grounds, was hard for teams to use as practical guidance, it rewrote the policy as a team effort. Instead of mandating one tool it offers choice within a bounded set of best-in-class tools: an organization-wide assistant for general productivity, and on the developer desktop IDE plug-ins, Cursor or terminal assistants, used with local open models or frontier models under data-privacy agreements. Code review stays human and engineers remain responsible for the commit.

Sourcegraph’s security team receives every new vulnerability report from its bug bounty platform through a signed webhook to a small service, which verifies the signature and passes the report to the Deep Search agent to investigate it in the relevant codebase. The agent judges validity, severity and affected code, suggests fixes, and posts the result as a team-only comment on the report; the engineer finds the investigation done in the morning and can fork the conversation to ask follow-ups.

In its company-wide Build Week, Buffer lets 72 people set aside regular work for a week to build internal tools. Earlier years paired people into teams on pre-selected projects; in the third edition everyone chose their own project, built with AI, and drew on experts across the company when needed. The week produced 112 completed projects, 1,083 merged pull requests across 26 repositories, and six people merging their first pull request.