An AI teammate that builds itself: how we wired Avery into our company’s DNA

We built an AI teammate called Avery into the way our engineering team works.

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.

Why it's different from a coding assistant

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, and what sets it apart isn't raw capability - it's context. It's wired into every system we actually run on:

• It runs on real Mac hardware. Avery compiles and runs the Extra iOS app, logs into test accounts, and records the fix working. An agent on a Linux box can read Swift, but it can't run it or show you the result.
• It debugs from production, not guesses. Avery pulls our production GCP logs, databases, analytics, and every bug channel together, so a vague report becomes a root cause backed by data.
• It owns the whole PR lifecycle. From one Slack message, Avery spins up an isolated worktree, implements, tests, opens the PR, links the ticket, and tracks it back to the thread - then merges and cleans up when asked. Each task gets its own worktree, so several engineers can hand Avery work in parallel without stepping on each other.
• It's versioned with the company. Avery's skills, scheduled jobs, and team conventions live in the repo next to the code. They improve through the same pull requests as everything else - and Avery can open those PRs against its own skills, reviewed like any other change. There is no no-code dashboard and no self-editing outside of review. Avery builds itself.

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.

Avery proactively fixes bugs

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.

Fixing on the go

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. Below is an example of how we use Avery on mobile.

Implementing new features

One of Avery's superpowers is implementing new features, with live verification through screenshots and videos. Here's Steven walking through a common workflow engineers use.

Data analytics

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.

Skills that run on their own

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.

The result is concrete: bugs are triaged twice a day without a human doing the first pass, every engineer starts the day with a personalized focus brief, production errors are summarized every few hours, and release readiness is reviewed weekly without anyone manually pulling the data. 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.

Privacy and guardrails

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.

Native Mac and iOS apps

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.

The takeaway

Avery does not decide product direction - how the product should feel, what users should trust, which tradeoffs are worth making. That still comes from the team, and Avery gives them more room to focus on it.

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, and that's why we built with Avery.

We're also hiring engineers who want to work this way.

If you're building something like this, or would like to embed Avery at your company, we'd love to chat. Reach out to the team at avery@buildforever.com.

Oh, and if you're wondering: Avery is a real person too - the incredibly cute three-year-old of one of our engineers, Luke. The human Avery is still working on his first PR, but he's got time.