After the code is written, what does AI take over: YC W26 Operations/Testing/Automation/Agent Infrastructure 22 companies fully deconstructed

Written by: Lang Hanwei Will

This is the fifth article analyzing the YC W26 series. The previous one covered AI programming tools and “Claude Code for X” (12 companies). This time, we look at the other half of the development chain—everything after code is written: operations, testing, workflow automation, and infrastructure for Agent development, totaling 22 companies.

Code is just the beginning

In the last article, we discussed how AI is changing “coding.” But coding is only part of software development—after writing code, there’s deployment, operations, monitoring, testing, bug fixing, and workflow automation, each step requiring human effort.

The 22 companies in YC W26 are working to delegate every step after code completion to AI agents.

Early morning production alerts? IncidentFox automatically checks logs, identifies root causes, and prepares fix scripts while you sleep. You just review and approve when you wake up. User reports a bug? Lucent monitors session replays 24/7, often discovering issues before users do. Need to automate approval workflows in Excel? Bubble Lab can do it in one sentence.

These 22 companies can be divided into four groups: AI Operations/SRE (5 companies), AI Testing/QA (2 companies), AI Workflow Automation (7 companies), and Infrastructure for Agent Development (8 companies).

Here are four key takeaways:

1. IncidentFox has the most complete product among these 22. Two former Roblox engineers (supporting infrastructure for over 100 million daily active users), open source, with 300+ pre-built integrations, deploys in under a day. Its core differentiation isn’t just “AI log analysis” (which anyone can do), but “automatic discovery of your tech stack and auto-generation of integrations”—eliminating the most painful onboarding work.

2. AI Operations (5 companies) and AI Testing (2 companies) aim to transform “code quality assurance” from labor-intensive to agent-intensive. Traditional operations, bug investigation, and regression testing require significant engineer time. These 7 companies bet that agents can detect issues faster, pinpoint root causes quicker, and operate 24/7 without rest.

3. Workflow automation (7 companies) is the most diverse and widely used group. Their common goal is “enabling non-coders to automate work with AI”—RamAIn uses computer vision to operate any software, Bubble Lab creates automation flows with one sentence, Jinba automates enterprise workflows via chat. They target knowledge workers, not developers.

4. Infrastructure for Agent Development (8 companies) is the most “meta” group—building tools for those creating agents. Emdash offers an open-source environment for agent development, Overshoot provides an AI visual application platform, Glue offers a canvas for designing agent interfaces. Their logic is similar to the “agent economic infrastructure” discussed in the fintech article: as agent numbers explode, tools for building agents become a necessity.

Subsector 1: AI Operations/SRE—IncidentFox, Mendral, Corelayer, Sonarly, Lucent

Five companies focus on different aspects of replacing operations engineers with AI.

IncidentFox

Official website:

AI SRE agent—automates triage, investigation, and repair of production incidents, integrated within Slack.

Core data: open source (Apache 2.0), 420+ GitHub stars, 300+ pre-built integrations, supports Kubernetes/AWS/Grafana/Prometheus/Datadog/PagerDuty/GitHub.

Business model: core open source + enterprise version (security sandbox, credential proxy, multi-team management). Deployment in under a day.

Team highlights: Jimmy Wei—former Roblox engineer (built social messaging features supporting 100+ million DAU), previously at Meta FAIR doing multi-party conversational AI research, Cornell CS. Long Yi—former Roblox infrastructure engineer (database infrastructure supporting 100+ million DAU). Two founders—one builds AI, the other handles operations—complementary skills.

Competitors/Risks: PagerDuty, Incident.io (raised over $50M), Datadog, ServiceNow are expanding into AI ops. IncidentFox’s differentiation is “auto-generation of integrations”—other tools require weeks of manual setup; IncidentFox analyzes your codebase and incident history to generate integrations automatically.

Other highlights: SOC 2 compliant. Each investigation runs in isolated containers; agents cannot see raw keys. Also released Claude Code plugin for individual developers.

IncidentFox’s core insight: the failure of AI ops tools isn’t due to weak models but insufficient integration. Your payment team uses a custom Kafka pipeline, your infrastructure team has a custom deployment system, your ML team has custom model services—all these prevent general AI tools from plugging in. IncidentFox analyzes your codebase and incident history to automatically discover missing integrations and generate them—humans only need to approve.

Chris Lu described IncidentFox as “AI SRE engineers independently fixing production incidents.” This positioning is both a boon and a threat to traditional ops engineers.

Mendral (90K monthly visits) develops AI DevOps engineers—focused on daily operations like continuous integration, deployment management, environment configuration, etc.

Corelayer (40K monthly visits) builds “data-driven AI on-call engineers”—emphasizing data-driven debugging, automatically correlating metrics and logs rather than guessing.

Sonarly (20K monthly visits) creates AI engineers for production alerts—classifies, deduplicates, and correlates alerts to identify critical issues needing human attention.

Lucent (160K monthly visits) automates session replay analysis to detect bugs—not from code but from user experience. AI watches user sessions 24/7, finds lag, errors, anomalies, and automatically creates bug tickets in Slack and Linear with full reproduction context.

Founder Alisa Rae’s story is worth noting: Australian, previously bootstrapped and sold an edtech company, was the second employee at MagicBrief (acquired by Canva), worked on rich text editors at Atlassian. Rejected YC on her first try, advised to find a co-founder, but she persisted solo, raised $2M seed, and was accepted on her second attempt. Used by over 30 YC companies; founders report: “Discovered 7 previously unknown bugs in the first week” and “broke even in the first week.” Since 94% of users encountering bugs don’t report them, they just churn—this is Lucent’s raison d’être.

The common logic among these five: most of an ops engineer’s time isn’t spent fixing problems but finding them—correlating signals from dozens of monitoring systems, sifting logs, checking recent deployments—this investigation accounts for about 80% of repair time. AI agents can query all data sources simultaneously, correlate in seconds, reducing “problem hunting” from hours to minutes.

Subsector 2: AI Testing/QA—Canary, Ashr

Two companies focus on AI testing.

Canary

Official website:

“First AI QA engineer that understands your codebase.” Key phrase: “understands codebase”—not a generic testing tool, but one that comprehends your logic and generates targeted test cases. Traditional AI test generators often produce disconnected test cases.

Ashr

Automates multimodal testing with agents—“multimodal” means testing not just text interfaces but also images, videos, speech, etc. As AI applications increasingly use multimodal inputs/outputs, testing tools must evolve accordingly.

Subsector 3: AI Workflow Automation—RamAIn, Bubble Lab, Jinba, Ressl AI, EigenPal, Carson, Crow

This is the broadest group—targeting all knowledge workers, not just developers.

RamAIn

Official website:

“World’s fastest computer using agents”—teaches AI to operate your computer like a human, moving data between browsers and desktop apps.

Core data: 35K monthly visits, used by procurement, insurance, healthcare, finance teams. Deployment in days.

Team highlights: Two IIT Delhi students—CEO Shourya previously at McKinsey on enterprise AI projects, founded Genoshi (AI studio, bootstrap to six figures), FIDE chess player (2118 rating), represented India in 17 countries.

Business model: enterprise-level—automates data transfer between legacy systems, desktop apps, and web portals. Targets ERP, insurance brokers, hospitals, finance teams.

Competitors/Risks: Anthropic’s Computer Use, OpenAI’s Operator—biggest threats. RamAIn’s differentiation: “pre-trained on specific interfaces”—general CUA (screenshot→vision model→decision→repeat) is costly, slow, unstable. RamAIn learns your interface first, then automates. Also features “self-healing”—UI changes won’t break it, unlike traditional RPA.

Bubble Lab (19K monthly visits) automates workflows with one prompt. Converts repetitive tasks into automation flows with a single sentence—simpler than Zapier, which requires configuring triggers and steps.

Jinba (17K monthly visits) automates enterprise workflows via chat—trigger approvals, data flows, system integrations through chat interfaces.

Ressl AI (17K monthly visits) configures ERP/CRM systems—uses AI agents to handle complex setup and customization after Salesforce or SAP.

EigenPal (9K monthly visits) manages enterprise AI document workflows. Carson (already detailed in OpenClaw) creates desktop AI workspaces. Crow (25K monthly visits) adds an AI chat layer to SaaS products, enabling users to control apps via chat without learning interfaces.

The common logic: AI programming lowers the barrier to “coding,” but most work doesn’t require coding—it’s about connecting existing tools and automating repetitive workflows. These companies focus on “no-code automation.”

Subsector 4: Infrastructure for Agent Development—Emdash, Overshoot, Cardboard, Glue, Sila, Valgo, SideKit, Wideframe

Tools for building agents.

Emdash (23K monthly visits): open-source agent development environment—over 60K downloads, 2430 GitHub stars. Supports parallel running of multiple coding agents, compatible with any model provider. Similar to the previous “1 code” approach but emphasizes open source and model-agnosticism.

Official website:

Overshoot (16K monthly visits): AI visual application platform—helps developers build and run AI visual apps. As multimodal models grow, “visual AI” is a rapidly expanding category.

Cardboard (7K monthly visits): AI video editing tool—automatically cuts, stitches, adds subtitles and effects. Traditionally requires professional skills and expensive software; Cardboard aims to lower the barrier to “tell the agent what you want.”

Glue: creates interfaces for agents—designs front-end panels when needed. As more agents require visual interfaces, demand will grow.

Sila: infrastructure for agent communication—enables message passing between multiple agents working together.

Valgo (3K monthly visits): autonomous system security verification. SideKit (2K monthly visits): one-stop solution for mobile app deployment (rare non-AI company in this batch). Wideframe: AI video editing collaborator.

Looking at all 22 together

Some observations:

First, AI Ops (5 companies) is the most mature group. IncidentFox is open source, with 300+ integrations and SOC 2 compliance. This isn’t accidental—operations is one of the easiest AI use cases to demonstrate value: repair times drop from hours to minutes, measurable directly.

Second, workflow automation (7 companies) faces competition not from each other but from existing tools like Zapier, Make, n8n. AI makes these tools smarter, but they are also rapidly adding AI features. To survive against Zapier (valuation over $5 billion), these startups need narrow, differentiated niches.

Third, infrastructure for agent development (8 companies) is a long-term bet. Currently, agent counts are low, and infrastructure value isn’t obvious. But if the agent economy explodes (as seen with Sponge enabling agents to open bank accounts in fintech), the toolchain for building agents could become a cloud-infrastructure-level opportunity.

Fourth, all 22 are B2B. Like all articles in this series—YC W26 is a thoroughly B2B batch. These AI tools are sold to enterprises and developers, not consumers.

Implications for Chinese teams

First, AI Ops demand is huge in China. Chinese internet giants (ByteDance, Alibaba, Tencent, Meituan) have large operations, but their AI-driven tools are less advanced than in the US. Domestic monitoring systems (Alibaba’s ARMS, ByteDance’s APMPlus) haven’t pushed AI features like Datadog. Building a “Chinese IncidentFox”—integrating with local monitoring systems, supporting Chinese logs, understanding domestic tech stacks—has market potential.

Second, workflow automation has a special scene in China—DingTalk and Feishu. These platforms are core work portals, but their automation capabilities are still basic. Developing “AI workflow automation in DingTalk/Feishu” (like Jinba for Slack) could gain quick adoption over building a new platform from scratch.

Third, agent development tools are still absent in China. The US has Emdash, Glue, Sila; China lacks equivalent tools. As more Chinese developers enter agent creation, this market will open.

Key takeaways

1. The core bottleneck in AI Ops isn’t model capability but integration depth. IncidentFox’s “auto-generate integrations” approach is worth emulating—no matter how smart your AI, if it can’t connect to customer systems, it’s useless.

2. “No-code automation” is becoming an independent category. The traffic of RamAIn, Bubble Lab, Crow shows real demand. These tools target all knowledge workers, not just developers—this market is ten times larger than developer tools.

3. Infrastructure for agent development is a “long-term correct but short-term unprofitable” direction. Like cloud infrastructure in 2010—initially seen as overkill, but later became the most profitable layer. Agent infrastructure may follow the same path.

4. Combining these 22 companies with the previous 12, the DevTools track totals 34 companies—by far the largest in W26. This reflects a fundamental truth: AI first changes “how software is built,” then transforms other industries. Developer tools are AI’s “home base.”