Pattern 01 · Outcomes, not tickets
“What would it take to leave Cloudflare Stream?” Answered and shipped in one session.
- Devin scoped production (24 live releases of 2,563 videos actually mattered), migrated 517K objects / 201 GB to R2, built a per-release flag and flipped all 24.
- Along the way it found and fixed two latent production bugs the hosted provider could never fix: a timestamp defect that made deep-seek loop forever on 5 long shows (163,899 segments rewritten and re-verified at the edge), and replays opening at the live edge instead of 0:00.
- Then it shipped resume-on-reload, verified in a production browser, and escalated rather than guessed when a flip touched a shared deployment.
What it shows: Devin owns the whole arc: discovery, migration, flag, verification, follow-on bugs. I made four decisions.
Pattern 02 · Measure, then decide
The portrait clipper: a general-purpose model lost to a $0 local detector, by measurement.
- The first version used a frontier vision model to reframe concert footage to 9:16. Testing on real footage showed it lost the singer for 22 of 30 seconds.
- My push: “We're using a general-purpose model for something super specific. Something local should be explored too.”
- Devin benchmarked 7 gateway models on 180 hand-labeled frames plus YOLOv8n/YuNet locally: mean column error, bias, wide-shot accuracy, latency, cost.
- Local YOLOv8n won outright: 0.50 mean error, 99% wide accuracy, ~1.4s, $0, versus the best hosted model at 0.55 / 61% / ~11s / $0.028 per frame. The gateway was removed from the render path.
What it shows: Taste asked the question; the agent built the harness, ran the bake-off and reported the table honestly: “0 shots worse, 7 better.”
Pattern 03 · Evals as the spec
Live song recognition: evaluated on 450 real shows before it shipped.
- Devin built a leave-shows-out eval harness, and first reported honestly that it couldn't evaluate: only 3 labeled shows existed in the dev data.
- Re-run on production labels: 450 shows, 6,157 labeled tracks, 309 songs.
- Result: CLEWS embeddings 0.752 window accuracy vs 0.222 for CLAP; setlist recall 0.978 vs 0.829; boundary error 21.7s vs 87.5s.
- My next question, “think we can identify a mid-Type-II jam?”, became three measurable signatures out of the same harness.
What it shows: Same culture everywhere: agent-behavior evals run in CI with persona scenarios, and found a real product bug without touching product code.
Pattern 04 · Parallel waves → one PR
A new product from an empty repo, in numbered waves of 4–8 parallel Devins.
- “Let's create a new repo for an app called Deep… Go! Also put together a set of parallel subagents.”
- Each wave: 4–8 child sessions, each on its own machine with its own PR (auth, sensors, grants, MCP server, web dashboard, iOS, Android, Watch, evals), consolidated into one integration PR per wave.
- My entire input across 15 waves reads like: “merged, do next wave” · “merge wave 3 into one and reparent to main” · “conflict on 6” · “Is the Expo app bootstrapped yet?!”
- While consolidating, Devin found and fixed a production pagination bug nobody asked about, and flagged the privacy calls it had made on my behalf.
What it shows: Across the org, 67 parent sessions spawned 469 children. Merge conflicts are now resolved unprompted, a learned behavior stored in Devin's memory.
Pattern 05 · The agent runs the night
Night Mode: an automation Devin designed for itself.
- “Help me generate a new automation” → Devin drafted the prompt: fire every evening, nudge stuck sessions, audit merges / Sentry / PostHog / builds, spawn up to 5 overnight sessions, DM a Slack report.
- Hard guardrails: no merges, no production deploys, no user emails, every assumption logged as NIGHT MODE ASSUMPTION.
- A read-only Morning Brief at 5:47am triages every PR as READY / NEEDS YOU / BLOCKED: a 10-minute review over coffee.
- Real catches: an importer calling a mutation deleted the day before (~48 Sentry errors a day) found and fixed unattended; a 2.9× frame-stall anomaly; Revyl device-test drift; a red E2E suite on main.
What it shows: Nightly memory-refinement sessions prune and update Devin's long-term notes about how I like to work; new device-testing workflows are written back into the repo as skills.
Pattern 06 · Ops, not just code
Show night: a live broadcast with zero human ops.
- Delivery costs under $1 per show on the MediaLive → R2 origin, versus roughly $180–$1,800 per show on hosted providers at 10K viewers, with a 4K HEVC/HDR ladder none of them offered.
- Devin built and drilled the dual-pipeline channel (real test channel created, exercised, deleted), explained the $/hr tradeoffs in plain language, then pointed it at that night's show.
- When the truck's SRT feed “didn't connect”, Devin watched the R2 writes, confirmed real video (“guitarist under orange lights, 1080p24”), attached the feed to the live session, diagnosed the failure mode and stood up a $2/hr test lab for the second feed.
- Later it traced a 52-second per-rendition drift to the encoder's own logs and shipped a layered fix. When asked about cost, it priced the alternative and recommended against it.
What it shows: The same loop that writes code runs production: watch, verify, act, report.