Adopting AI Without Adopting Tech Debt (April 2026)

This is the blog adaptation of a talk I gave at the Austin CTO Club in January 2026.

TL;DR: Minimize AI-era tech debt through context window management (Skills, markdown, monorepos), smaller teams with clear system boundaries, and open-source tooling for MCP Apps, ChatGPT Apps, and Claude Connectors.

Hi, I’m Abe. I recently sold my company, Trigo, where I was CTO building AI in property tech. I’ve since been building sunpeak, the open-source MCP App framework for building and testing MCP Apps, ChatGPT Apps, and Claude Connectors.

Software development looks very different today, and I no longer think about tech debt the same way. I group my approach to minimizing AI-era tech debt into three areas: tech, people, and how we ship AI-native software.

---

Tech

Coding Agents

Ground zero for AI tech discussions. Leading coding agents include Claude Code, OpenAI Codex, Cursor, and Gemini Code Assist. Claude Code now offers up to 1M tokens of context on Opus 4.6, which means entire codebases fit in a single conversation.

Unlike humans, agents don’t struggle with deadlines or extra work. The real driver for tech debt in an agentic world is misalignment. Keeping all developer agents rowing in the right direction is key, and the way to do that is through context window management:

• Context windows range from 200k tokens to 1M tokens depending on the tool, roughly 500 to 2,500 pages of text
• We need to fit everything about our company, policies, vision, market, product, organization, team, project, personal preference, feature specification, and current conversation in this window
• We can host markdown from each relevant team and level of the organization hierarchy to keep all agents aligned over time
  • Agent Skills are one way to do this, implemented as directories with SKILLS.md files
  • IT maintains company policies and SOC instructions, division leaders maintain architecture and strategy docs, team leads maintain project-level context
• Monorepos remain the simplest way to keep context unified across related systems

Ancillary Tech

A 10x increase in coding speed requires a 10x increase in everything around it:

• Code review: CodeRabbit, GitHub Copilot code review
• Testing: coding agents do this well, and specialized frameworks matter for MCP Apps where you need to replicate host runtimes like ChatGPT and Claude (sunpeak’s testing framework handles this)
• CI/CD and infra: anything defined as code, agents handle well
• Security: Semgrep, Snyk, Dependabot with cooldowns

---

People

People are far more productive, but our human context windows haven’t changed in 200,000 years. To stay on the same page, teams need to be smaller.

• Smaller teams with tighter coordination
• Larger ownership per engineer

With empowered individuals and small teams, clear boundaries around systems matter more than ever to maximize autonomy and speed. If Amazon made the switch to a Service Oriented Architecture at 7,000 employees, that number is much lower today.

Additional tactics to keep people coordinated with AI:

• Agentic access to ticketing and SDLC systems via MCP (Linear, Jira, and GitHub all have MCP servers now)
• Dedicated early adopter(s) who champion tooling and share patterns with the team
• Clear product processes around specs, reviews, and launches
• Shared context documentation that both humans and agents reference

---

Shipping AI-Native Software

The proliferation of agents has come with new types of software that agents can use. Beyond Skills mentioned above, we builders are being tasked with building MCP servers, MCP Apps, ChatGPT Apps, and Claude Connectors.

MCP

Model Context Protocol (MCP) lets agents interact with our systems. If Skills are the static websites of the agent internet, MCP servers are the dynamic websites.

MCP moved from Anthropic to the Linux Foundation’s Agentic AI Foundation (AAIF) in late 2025, with platinum members including AWS, Anthropic, Google, Microsoft, and OpenAI. The 2026 roadmap focuses on transport scalability (stateless Streamable HTTP), enterprise readiness (audit trails, SSO-integrated auth, gateway behavior), and agent communication. There are now over 10,000 published MCP servers.

MCP will continue to change, but the rate of change is slowing as it matures under multi-stakeholder governance. Using open-source tooling to abstract the protocol layer remains the best way to reduce the tech debt surface area.

MCP Apps (ChatGPT Apps and Claude Connectors)

MCP Apps let anyone embed interactive, visual apps in AI conversations. ChatGPT Apps reach ChatGPT’s 800M+ users, while Claude Connectors bring data access and interactive UI to Claude across desktop, web, and mobile.

MCP Apps became an official MCP extension in January 2026. They’re now supported by ChatGPT, Claude, VS Code Copilot, Goose, and other MCP-compatible hosts. OpenAI launched the ChatGPT App Directory for third-party developers, and Anthropic expanded Claude Connectors with enterprise integrations for Google Workspace, Microsoft 365, and more.

Because these protocols sit on top of MCP and include host-specific extensions, they will keep evolving. A framework that handles the host differences and stays current with protocol changes saves real maintenance time. sunpeak does this: you write one MCP App codebase and it works across ChatGPT, Claude, and other hosts.

Testing MCP Apps Without Burning Credits

One of the biggest sources of friction (and cost) when building MCP Apps is the manual testing loop. Every code change means re-deploying, refreshing the host, and waiting for tool calls, often on a paid account burning API credits.

The sunpeak Inspector replicates the ChatGPT and Claude runtimes locally, so you can develop and test without any host account. The sunpeak testing framework adds automated testing on top: unit tests, E2E tests with Playwright, visual regression tests, and multi-model LLM evals, all running locally and in CI/CD. You can get started with the quickstart guide.

---