599 words
3 minutes

15 from Merge Debt: The 2026 Bill Your Parallel AI Agents Are Running Up

The landing order: parallel agents count once their work is in main, shaped and in a chosen order

By Docker Captain · CNCF Ambassador · IBM Champion
Dark monolithic cube in a deep void, three wide cyan beams pouring into its top while a single thin thread of light escapes from its base beside a small red indicator, a visual metaphor for merge debt, parallel agent work forced through a single landing point

Worktrees parallelize the work. Main is a single sequence. The gap between code agents have finished and code that has landed, reviewed and reconciled, is merge debt, and it grows in front of the one gate that did not scale: human review.

The controls below need no vendor claim. They make the debt visible and the repayment cheap enough to happen daily.

  1. 1

    Give every agent its own worktree

    Each agent edits its own copy of the tree, so nothing collides while the work is in flight. It is the right primitive for the work, and it moves the problem to the landing.

  2. 2

    Check what each worktree touches before you spawn another

    Two agents aimed at the same module is a collision you scheduled. A commit graph that shows every active worktree and its working changes in one view (GitLens 18 does) makes it a ten-second look.

  3. 3

    Give the landing order an owner

    When several branches finish together, the sequence is an engineering decision: smallest blast radius first, or the hottest file first, chosen, not left to whoever presses merge first. Owner: the engineer or lead who owns the target branch.

  4. 4

    Cap diff width per agent task

    Cap agent tasks the way you cap pull request size for people, and split before landing, not after the review stalls. Every downstream cost (rebase, review, revert) grows with the width.

  5. 5

    Make commit shaping the entry fee to main

    Machine output does not enter history as one blob called "update payments". Reviewable commits with meaningful messages, shaped by hand or by a tool such as Commit Composer (part of GitKraken AI, a paid plan), are required before review.

  6. 6

    Plot the debt next to velocity

    Two numbers per team on the same dashboard: time from agent-done to merged in main (the queue depth) and the share of agent-authored code reworked within two weeks (the quality echo). Owner: the engineering manager.

Three questions for the next engineering leadership meeting

  1. What is our median time from agent finished to merged in main, and is it growing?
  2. Who decided the landing order of the last three parallel branches that touched the same service, and was it decided at all?
  3. What share of AI-authored code merged last quarter was rewritten within two weeks?

What it rests on

The study behind the perception gap, and the tools the steps name. A measured run of one worktree per agent comes with the worktrees post on 2 December.

Generating is parallel. Landing is serial, and someone should choose the order.

The long version is the post Merge Debt. METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, 10 July 2025; GitKraken, GitLens 18.

Vladimir Mikhalev

Docker CaptainCNCF AmbassadorIBM Champion

The Verdict — production-tested analysis on YouTube.

The Verdict

Inconvenient truths about shipping in the AI era

Container security, platform engineering, and the agentic shift — tested in production, argued without the hype. The verdict reaches your inbox the moment there's one worth sending.

Related Posts

Same category
  1. 1
    Merge Debt: The 2026 Bill Your Parallel AI Agents Are Running Up
    Opinion & Culture · Three agents finish in parallel. Main takes one landing at a time. METR's perception gap, the rebase tax, and why Merge Debt is the number nobody plots.
  2. 2
    Your Knowledge Is a Depreciating Asset. Judgment Compounds.
    Opinion & Culture · AI made reproducible knowledge free, so technical expertise is now a depreciating asset. Judgment is the one that compounds. Here is how to move your weight.
  3. 3
    The Senior Engineer Signal: The 2026 Risk Your Velocity Metrics Hide
    Opinion & Culture · Juniors get the biggest boost from AI; seniors trust it least. That split is your earliest read on engineering risk, and on the talent you're about to lose.
  4. 4
    Agent Sprawl: The 2026 Engineering Risk Your Auditor Hasn't Named Yet
    Opinion & Culture · Unknown numbers of AI coding agents run in parallel — no audit trail, no isolation, no per-team measurement. By 2026 that's an audit finding.

More Blueprints

All 15
  1. 01
    The blast radius assessment
    Signal, 23 February 2026 · Map what an AI agent can break, and how fast a person can undo it, before it gets a single key: six steps, a worksheet and the repositories that measure the same rules.
  2. 02
    The AI intake gate
    Signal, 12 March 2026 · Find every AI tool in use, ask for its backend's access rules instead of a questionnaire, and block what cannot use SSO and central logging, before the first prompt.
  3. 03
    Build or buy a no-code AI platform
    Signal, 20 January 2026 · When a no-code computer vision subscription is the right buy, when it is a slow trap, three questions before the pilot and four clauses before you sign.
  4. 04
    The AI-BOM admission gate
    Signal, 18 October 2026 · Write a signed AI-BOM for every model from its lock line, read it at admission, and refuse any model unless every licence it declares is on the site's list.
The landing order: parallel agents count once their work is in main, shaped and in a chosen order
https://heyvaldemar.com/blueprints/landing-order/
Author
Vladimir Mikhalev
Published
2026-11-15
License
CC BY-NC-SA 4.0

Discussion

Open thread on GitHub →