-- Views
October 30, 26
スライド概要
(文末に日本語あり)
The shortest engineering path from ideas to products
Speaker: Makoto Mizukami (Senior Field Engineer, CircleCI Japan)
Slides from the speaker session at the Antler Japan x CircleCI Meetup for Startups, held on October 26, 2026.
Turning an idea into a successful startup still comes down to four phases that AI hasn’t changed—Idea Verification, Customer Problem Fit, Problem Solution Fit, and Product Market Fit. This session focuses on the one phase AI has reshaped—building your MVP—and the two things that get in the way when you build with AI: keeping up forward speed, and fighting the “gravity backward” of regressions.
It brings together good-old engineering fundamentals (small, frequent tests; local validation; lightweight automation) and modern approaches with CircleCI and Chunk, to help you ship your MVP faster and more reliably.
Data sources (incl. CircleCI’s State of Software Development Report Q2 2026) are cited within the slides.
The shortest engineering path from ideas to products
登壇者:水上 誠(CircleCI Japan / シニア フィールド エンジニア)
2026年10月26日に開催された「Antler Japan x CircleCI Meetup for Startups」のスピーカーセッション資料です。
アイデアを「成功するスタートアップ」へと最短距離で育てるには、AI時代でも変わらない4つのフェーズ(Idea Verification / Customer Problem Fit / Problem Solution Fit / Product Market Fit)があります。本セッションでは、そのうちAIが大きく変えた「MVP実装」に焦点を当て、AIを使って開発を速く進めるうえで直面する2つの課題——前進のスピードと、品質の揺り戻し(regression)——にどう向き合うかを解説します。
「good-old」なエンジニアリングの基本(小さく頻繁なテスト、ローカル検証、自動化)と、CircleCI・Chunkによる最新のアプローチを組み合わせ、MVPを速く・確実に届けるための実践的なヒントをまとめています。
※各種データは CircleCI「State of Software Development Report Q2 2026」ほか、スライド内に出典を記載しています。
開発の質と速度を底上げしよう! 世界中のエンジニアが愛用する CI/CD ツール「CircleCI」の公式アカウントから日本語で最新情報をお届けします。 最新イベント・ミートアップ情報はこちら: http://circleci.connpass.com
The shortest engineering path from ideas to products Makoto Mizukami Senior Field Engineer, CircleCI 1
Hi there! ● ● ● Makoto Mizukami Senior Field Engineer @ CircleCI Used to be a support engineer at Microsoft, a full-stack hands-on engineer at a legal-tech startup, and a people manager at CircleCI and Degica / KOMOJU 2
The shortest engineering path from ideas to products Makoto Mizukami Senior Field Engineer, CircleCI 3
The shortest engineering path from ideas to a successful startup products 4
tl;dr 5
The golden path for startups - it’s nothing new Four important phases to make ideas into a successful startup Idea verification Customer Problem Fit Problem Solution Fit Product Market Fit Crystallize your idea to the level that no one can mimic, no one can be more passionate than you, and no one can refuse to fall in love with it eventually. Interview possible buyers, put yourself in their shoes, and thoroughly understand their challenges and pains. Identify unique value proposition, define product customer journey, and design user experience. Develop a product / MVP that implements your ideas, test your hypotheses with the product in the market, and adjust your direction. 6
AI changed only 1 phase 7
The golden path for startups - it’s nothing new Four important phases to make ideas into a successful startup Idea verification Customer Problem Fit Problem Solution Fit Product Market Fit Crystallize your idea to the level that no one can mimic, no one can be more passionate than you, and no one can refuse to fall in love with it eventually. Interview possible buyers, put yourself in their shoes, and thoroughly understand their challenges and pains. Identify unique value proposition, define product customer journey, and design user experience. Develop a product / MVP that implements your ideas, test your hypotheses with the product in the market, and adjust your direction. Itʼs YOU being passionate about crazy ideas. Not AI. Itʼs YOU listening to and representing customers. Not AI. Itʼs YOU tailoring customer experience emotion. Not AI. It doesnʼt have to be you to code MVP. 8
CPF is yours. PSF is yours. CPF + PSF are your ideas. Coding MVP is not. 9
So concentrate on your ideas 10
The shortest engineering path from ideas to products Focusing on techniques to implement MVP efficiently Makoto Mizukami Senior Field Engineer, CircleCI 11
2 things that will bother you while implementing MVP with AI 12
Speed of forward progress Code generation by violently-tireless AI 01 13
AI enabled people to code a lot almost universally 14
… and generated a lot of work that are just piling up. It isnʼt making yourself running faster necessarily. 15
16
17
Source: DX, State of AI Impact in Engineering (data came from the report in Q2 2026 18
Source: DX, State of AI Impact in Engineering (data came from the report in Q2 2026 19
Actually some people managed to deliver more with AI 20
Thereʼs a huge gap in day-to-day throughput Source: CircleCI, State of Software Development Report Q2 2026 Pulse Report 21
While many orgs are struggling, the winners are widening the gap exponentially: p95 orgs are 9x faster on deliveries than the median Source: CircleCI, State of Software Development Report Q2 2026 Pulse Report 22
You need to deliver MVP faster than the “winners”. You cannot leave stuff piling up. 23
Gravity backward Regression, regression, regression! 02 24
MVP must reflect your ideas consistently 25
But AI is good at implementing something different inconsistently 26
Hint of that tendency: AI code tends to fail to pass tests Source: CircleCI, State of Software Development Report Q2 2026 Pulse Report 27
Then more back-and-forth was needed for corrections Source: CircleCI, State of Software Development Report Q2 2026 Pulse Report 28
Now we can understand why those people are spending a lot on AI p99 spending 11x what p90 spends p90 spending 54x what p50 spends Spend is up 3x in both p99 and p90 cohorts in 2026 A lot of spend to move forward and cancel the “gravity backwardˮ Can you afford it? Do you want to? 29
> things that will bother you while coding MVP 01 Speed of forward progress 02 Gravity backward How can you exploit the power of AI to move forward fast? How can you avoid the spoilers that push you backward? 30
What should we be doing then? 31
This was actually what the greatest engineers were doing. 32
What is the key essence of “good-oldˮ practices? ● ● ● ● Make a small investment in testing, and run tests small and repetitively ○ No need to be a fancy one, just a capability to run small tests (unlike time-consuming E2E tests) Test code locally before committing changes ○ Deliver only what you know. Do not bother your team with garbage you donʼt know what it is. Codify repetitive operations and automate it with minimal effort ○ At least “scriptizeˮ the testing procedure, building procedure, and deployment procedure ○ Then run tests, builds, and deployments automatically and autonomously ○ But never attempt to optimize the process / automation by yourself! Be aware that vibe coding can go in the opposite way! ○ It can generate a lot of untested code, without granularity or verifications ○ It can generate a lot of code at once, not granularly ○ It can fail to generalize what it built, making it harder to repeat at minimal cost 33
Then it should look like this holistically Change committed Outer loop System-managed) Inner loop Agent-driven) Debug Plan AutoBuild Release Note: This Release step is what your PMF experiment needs. Validate Local testing Code AutoDeploy AutoValidate Final Approval … then you want to run these loops as fast as possible! 34
The question is: How can it be made fast and effortlessly? 35
With CircleCI, these are available as an instant package Change committed Outer loop System-managed) Inner loop Agent-driven) Plan Debug Build Release Fast compute Validate Factory CI/CD Code Deploy Validate Smart Deployment Chunk sidecar Smarter testing Approval Microbuild 36
Chunk sidecars ● ● ● Mechanics for “shift-leftˮ: run validations as early as possible - while coding Coming with LLM configurations / skills, CLI, and microVMs mirroring CI environment 78 % faster than waiting for CI pipelines, 68 % cheaper than forcing agents to interact with CI 37
Smarter Testing - compilation of 4 features NEW PAID FEATURES Intelligent Test Selection Dynamic Test Splitting Auto-Rerun Failed Tests Test Analytics Dashboard Runs only tests impacted by your code changes via Test Impact Analysis. 100% confidence due to deterministic test selection. Distributes tests across parallel nodes dynamically during runtime using historical timing data, eliminating idle time and reducing wall-clock test time. Automatically retries failed tests at the individual test level without re-running the full suite. Test health data, skip rates, flaky trends, time saved to better understand your test suite performance. Coming Soon Save time (and costs) up to 80 % 38
Fast and efficient computation ● ● Adjust resource allocation: less CPU cores for single-threaded tasks, more memory for intensed jobs Modern “Gen2ˮ infrastructure - 40 % faster, 28 % cost saving 39
Smart deployments ● ● Make both build and deployment progress visible in one place - on CircleCI Manage version promotions and rollbacks from CircleCI - both manually and even automatically 40
The ultimate goal The demo presented at the SF Tech Week Workshop 41
Let’s reach PMF wisely, and as fast as possible. By applying good-old best practices and cutting-edge approaches 43
Exclusive offer for startups! ● ● ● CircleCI for Startups helps you start using Chunk sidecars and CircleCI overall 10,000 USD - up to 20,000 USD in free compute credits, for up to 24 months Free access to paid support (a 249 USD/month value), and gated features like machines with GPUs Visit https://circleci.com/startup-program/ OR http://antler.getproven.com/vendor/circleci-3 OR https://startups.aws.com/offers 44
Thank you. 45