K Koda Intelligence
DEEP DIVE DEEP DIVE · 15 October 2026DOCNO RECEIPT FILEDno checked claims on record for this article
FILED 15 OCTOBER 2026

Press Enter Tonight. Wake Up to Pull Requests. What Claude Code Can Do Now

Bun was ported from Zig to Rust with Claude Code: roughly 750,000 lines in eleven days, with 99.8% of the tests passing. Here is how the night shift works, and what it costs.

SPECIAL EDITION · 3 MIN READ · BY THE KODA EDITORIAL TEAM · OCT 15 2026
BUN PORT750,000NOT MEASUREDLINES OF RUST
FIRST COMMIT TO MERGE11NOT MEASUREDDAYS
TESTS PASSING99.8%NOT MEASUREDEXISTING SUITE
BUN PORT750,000LINES OF RUST FIRST COMMIT TO MERGE11DAYS TESTS PASSING99.8%EXISTING SUITE CODE REVIEW6 to 10MINUTES, MOST REVIEWS OPUS 5.5 VS OPUS 540%LESS TO RUN, ANTHROPIC REVIEWERS PER FILE2ON THE BUN PORT

You press Enter before you log off. By morning, the pull requests are waiting. That is the short version of what Claude Code can do in October 2026, and it is not a demo: Bun, the JavaScript runtime, was ported from Zig to Rust this way. We made a one-minute film of that night shift, set in a miniature city under the keyboard. You can watch it on this page, or as a YouTube Short.

One plan, hundreds of agents

The feature behind it is called dynamic workflows. You give Claude Code one large job. Claude writes its own orchestration plan, breaks the job into pieces, and runs those pieces across tens to hundreds of subagents in a single session. Results are checked before they are folded back in, and progress is saved as the run goes, so an interrupted job picks up where it stopped. You can ask for a workflow directly, or turn on the ultracode setting and let Claude decide when a task needs one.

READ THE NUMBERS WITH CARE · OCTOBER 2026ANTHROPIC'S OWN PAGES · CHECKED OCT 15 2026BASE: NO CLAIM LEDGER FILED, 4 SHOWN

What each number does and does not say

Lines in the Bun port one team's result; Anthropic says it is not yet in production NOT MEASURED
750,000
Tests passing of Bun's existing suite, after the port NOT MEASURED
99.8%
Code Review time most reviews, Team and Enterprise only NOT MEASURED
6 to 10
Opus 5.5 running cost vs Opus 5 on typical workloads, by Anthropic's tests NOT MEASURED
40%

The Bun port, by the numbers

Anthropic's announcement gives the clearest example. Jarred Sumner used dynamic workflows to port Bun from Zig to Rust: roughly 750,000 lines of Rust, 11 days from first commit to merge, with 99.8% of the existing test suite passing. One workflow mapped the right Rust lifetime for every struct field. The next wrote every file as a behaviour-identical port, with hundreds of agents working in parallel and two reviewers on each file. A fix loop then drove the build and the tests until both ran clean. After the port landed, an overnight workflow went after unnecessary data copies and opened a pull request for each one.

Go to bed with a plan. Wake up to the pull requests.· THE LAST LINE OF THE FILM

Anthropic is careful to add that the port is not yet in production. Read the numbers as what one well-resourced team achieved, not as a promise for your codebase.

Three things to act on

SCALE
11

Days for a 750,000-line port

Big migrations become an overnight job. Pick one you have been postponing and scope a slice of it.

CONTROL
2

Reviewers on every file

Independent checking is built in. Keep a person as the final approver, like any other change.

COST
40%

Cheaper model, heavier runs

Opus 5.5 costs less to run, but a workflow uses far more usage than a normal session. Measure one small job first.

Checked before you see it

The part that interests me most is the checking. In a workflow, agents attack the problem from independent angles, other agents try to refute what they found, and the run keeps iterating until the answers converge. On Bun, every file had 2 reviewers. Nothing reaches you until it has survived someone whose job was to break it.

What else shipped this month

Code Review left research preview on October 15 for Team and Enterprise plans. Anthropic says most reviews now finish in 6 to 10 minutes, about 3x faster than in preview, and admins can choose a quick, standard or deep pass for each repository.

Mods arrived on October 1. A mod is a small TypeScript function that hooks into what Claude Code does: it can rewrite a prompt, block or retry a tool call, redact secrets, or add buttons and panels to the interface. You can write one yourself, or ask Claude to write it, install it and hot-reload it in your session. Even the built-in diff view now ships as a mod you can replace.

Underneath all of it sits Opus 5.5, released on September 22. Anthropic says it performs at the level of Fable 5.1 on most work and costs 40% less to run than Opus 5 on typical workloads.

Why a finance team should care

I run AI on real finance work, and I read this the way I read any operating model. The speed is the headline. The design is the story: independent reviewers, adversarial checks, and a person who approves the result at the end. That is segregation of duties, built into the tool. It is the same control logic we use for journal entries and payments, applied to code.

The catch

Workflows are not free. Anthropic says plainly that they can use substantially more tokens than a normal Claude Code session, and recommends starting on a scoped task to get a feel for the usage. Before you hand a migration to an overnight shift, run one small job, measure what it cost, and decide from there.

The bottom line

Go to bed with a plan. Wake up to the pull requests. Then review them like you would review anyone else's work.

Sources. Anthropic, Introducing dynamic workflows, checked on October 15, 2026. Anthropic, Code Review is now generally available, October 15, 2026. Anthropic, Customize Claude Code with mods, October 1, 2026. Anthropic, Introducing Claude Opus 5.5, September 22, 2026.

The miniature city in the film is an illustration, not footage of Claude Code or of Bun's repository.

DOJO · BEFORE YOU HAND OVER A NIGHT SHIFT

Three steps to try this week

  1. Scope one job. Ask Claude Code to create a workflow for a small, well-tested part of your codebase.
  2. Read the checks. See what the reviewing agents flagged before you read the final pull request.
  3. Count the cost. Note the usage of that one run before you schedule a bigger one.
Train the full skill in The Dojo
THE BOTTOM LINE

The speed is the headline. The checking is the story.

Claude Code can now plan, split and verify large jobs across many agents while you sleep. Treat its pull requests like anyone else's: reviewed, approved, and measured for cost.

WATCH · VISUAL NARRATIVEAnimated breakdown · ~1 min
PLAY · YOUTUBE
Filed underAgentsDeep Dive15 October 2026
Browse the Deep Dive archive

Get the morning Signal

204 editions so far, one a day. Unsubscribe anytime.