Avery is an AI engineering teammate wired into the systems and tools BuildForever's team uses every day — and it now writes nearly a third of the code the team ships to build Extra every week.
Last week, Avery wrote 31% of the code our five-engineer team shipped to production. But not just by being prompted - it identified the issues, diagnosed the root causes, generated the fixes, and carried them through to merged production code on its own. For the tiny team we are, that's a step-change in leverage: an always-on teammate turning bugs, alerts, and maintenance work into shipped code around the clock. What started as a side project is now woven into every part of BuildForever, and now lets anyone at the company ship code to production.
Avery lives in Slack, runs on its own Mac mini, and has access to the systems our team uses every day: GitHub for code, pull requests, and reviews; Linear for issues; email and WhatsApp for bug reports and community feedback; TestFlight and App Store Connect for crash reports and user reviews; Mixpanel for product analytics; and production logs in Cloud Logging to trace what actually happened.
You talk to Avery in Slack. Mention it in a thread and it spins up a persistent agent session tied to that thread, so the conversation and the work behind it stay linked. Follow-ups resume the same session instead of starting over, it carries a durable memory across threads, and it runs scheduled jobs on its own without waiting to be asked. That persistence puts Avery in the same family as persistent agents like Claude Code or Devin.
But Avery is not just a code-completion plugin - what sets it apart is its proactivity, memory and context. It's wired into every system we actually run on:
The next wave of AI engineering tools won't just live in the IDE. They'll be embedded in the company's DNA - the operating systems, tools, human conversations, logs, analytics, and rituals that turn code into product.
From a single user report, Avery digs through production logs to find the root cause, opens a PR with the fix, runs the app on its Mac mini to verify it, and reports back in the same Slack thread.
Avery always does the first pass, freeing up the team to focus on other critical work.
Avery also set up automatic iOS crash reporting with BugSnag. A new crash shows up, Avery picks it up on its own, traces it to the root cause, opens a PR with the fix, and verifies it - without any engineer needing to be in the loop.
Avery also constantly watches our backend for trends in API errors. When a class of errors starts spiking, it alerts the team, traces the regression through the worker logs against a baseline, and puts up a fix - before it turns into a page.
When we're away from our computers, Avery on mobile has been a game changer. Any bug we hit while using Extra, we can hand straight to Avery and fix it on the go - which has 10x'd how fast we turn around user-facing fixes against our SLAs.
One of Avery's superpowers is implementing new features, with live verification through screenshots and videos. Here's Steven, co-founder of BuildForever, showing how Avery can build new features.
Because Avery lives in any Slack channel, the whole team can see how everyone else uses Avery, which turns each request into a chance to learn - you pick up new ways to prompt just by watching, and anyone can chime in mid-thread to nudge, correct, or extend what's happening. The result is the most multiplayer prompting we've seen: several people steering Avery toward a change together, in the open, instead of working in private one-on-one sessions.
Analytics is one of the clearer wins. Ask whether a new feature is being adopted, and Avery pulls the Mixpanel numbers, our backend session logs, finds where users drop off, updates the dashboard, and posts the answer in-thread - so product questions don't have to wait for someone to free up.
Avery doesn't only act when it's asked. Some of its most valuable work runs on a schedule, with nobody in the loop, posting results where the team already looks:
Engineering Daily Pulse - every weekday evening, a read on what shipped, what regressed, and what still needs attention.
Personal Morning Focus - each teammate gets an early brief: what's on their plate, what's waiting on them, what to focus on today.
Daily Bug Sweep - twice a day, sweeps every report channel into triaged, de-duplicated issues with the evidence already attached.
iOS & Web PR Review Queues - weekly queues that surface the riskiest changes still waiting on review.
Security Log Review & API Pentest - a daily pass over the security logs and an automated penetration test of our own APIs.
Error Digests - a Cloud Logging error digest every few hours, so regressions surface fast.
Release Report - a weekly readiness report for the next iOS ship.
Today Tab Quality Review - a weekly product-quality pass on the feed itself.
Because these skills live in the repo, each run holds the same standard as the last, and Avery sharpens them through PRs it can open itself.
Avery works with people's email and user reports, so its guardrails are explicit. It never posts a user's personal information or email contents into Slack, even when the whole team is in the thread.
When a bug report comes in, Avery investigates it and proposes a fix working only from what that user has chosen to share with us. The same privacy bar we hold the product to applies to the agent.
Slack is Avery's home base, but it isn't the only way to reach it. Avery also runs as native Mac and iOS apps, so the same teammate is a click away on a laptop or a phone.
That changes onboarding. When someone new joins, they don't work through a setup doc - they install Avery, and it walks them through getting set up in seconds: the tools to connect, the accounts to provision, where the team already works. The first thing a new hire meets is the teammate that already knows how the company runs.
Avery only became useful after a few weeks of very specific tuning. Early on it would stop at diagnosis, over-trust stale context, open PRs without enough proof, reply in the wrong Slack context, or patch symptoms instead of root causes. So we turned every failure into operating instructions: use fresh worktrees, inspect thread history, ground claims in production evidence, show screenshots for UI fixes, run focused tests, write session logs, and leave risky merges to humans. These days Avery does a lot of that work on itself - we use Avery to build Avery, drafting its own skills, fixing its own bugs, and shipping its own improvements. That tuning is also what makes Avery feel different: it adapts quickly, has personality, and learns how the company actually operates so it can weave into the existing systems instead of becoming another tool to manage. We still override it on taste, architecture, privacy, and judgment-heavy calls.
The lesson for us is that the next wave of AI engineering tools won't just live in the IDE. They'll be embedded in the company's DNA - the operating systems, tools, human conversations, logs, analytics, and rituals that turn code into product, which is far more useful than dropping a generic agent into a workflow.
If you're building something like this, or would like to embed Avery at your company, we'd love to chat. Just email Avery at avery@buildforever.com
We're also hiring engineers who want to work this way. Reach out to careers@buildforever.com
Oh, and if you're wondering: Avery is a real person too - the incredibly cute one-and-a-half-year-old of one of our engineers, Luke. The human Avery is still working on his first PR, but he's got time.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。