Catalyst Agent Skills: Teaching AI agents how to build on Catalyst

Zoho's Catalyst Agent Skills, an open-source set of model-facing documentation guides for its full-stack cloud platform, enabled the smallest AI model to complete 92% of an end-to-end full-stack app build task, compared with 25% for the strongest AI model without platform context, according to the company's September 21, 2026 blog post. The guides cover services including Data Store, Functions, Authentication, Job Scheduling, Cache, Stratus object storage, and AppSail, and load only the single relevant guide per request to conserve tokens. Zoho frames the release as closing the gap where models repeatedly burn tokens re-establishing platform context across sessions.

- HOME https://catalyst.zoho.com/blog - Catalyst Agent Skills: Teaching AI agents how to build on Catalyst Catalyst Agent Skills: Teaching AI agents how to build on Catalyst - Last Updated : September 21, 2026 - 47 Views - 7 Min Read Given the same task of building a full-stack app end-to-end on Catalyst, Zoho's agent-ready full-stack cloud platform, the strongest AI model without platform context finished 25% of it, and the smallest model with Agent Skills finished 92%. This blog is about why. LLMs are no longer an experiment. It is becoming as embedded in how we work as search once was: in our workflows, our decisions, and our development cycles. The question is no longer whether to use it; it's how well you do. Tokens are not free Here is the one technical idea worth understanding, whatever your role. Everything you type to a model, and everything it replies, is broken into tokens, which are, roughly, pieces of words. A model has a finite working memory measured in tokens. Every token spent on context is a token no longer available for reasoning. That has a direct consequence: the quality of what you put in determines the quality of what comes out, and the efficiency of what you put in determines how much room the model has to think. A precise prompt with relevant context beats a vague prompt padded with everything you could find. For a single question, this is a nice optimization. For teams building real software with AI in the loop, where a model may work through hundreds of steps on your behalf, token efficiency becomes architectural. And if you own a budget rather than a codebase, tokens are the line item: they are what you pay for, and a model that runs out of working room halfway through a task starts forgetting what it was doing, which you then pay to re-explain. The problem we set out to solve Every team that builds on a platform pays a recurring cost: learning it from zero. It comes due each time a new person joins, each time someone moves to a different project, and now each time an AI model starts a fresh session with no memory of the last one. The first two have always been part of running a team. The third happens dozens of times a day. For a developer, it looks like this. They ask an AI model to help them build a full-stack application on Catalyst. The model knows about Catalyst: generic concepts and patterns, plus whatever fragments of Catalyst happened to be in its training data. So it produces code that looks plausible and often doesn't account for how Catalyst actually behaves. The developer spots the mistake and corrects it. The model adjusts. The next session, the model has forgotten, and the cycle restarts. Tokens are burned re-establishing context that should have been there from the first prompt. This is the gap Catalyst Agent Skills https://github.com/catalystbyzoho/agent-skills exists to close. What Agent Skills are The simplest way we can put it is that Agent Skills are documentation written for models, not for humans to read. Catalyst Agent Skills are a curated, open-source set of these model-facing guides, one for each service across the platform's full stack: Data Store, Functions, Authentication, Job Scheduling, Cache, Stratus object storage , AppSail, and more. They are organized so that a model doesn't have to read everything to answer anything. A lightweight entry point recognizes what the developer is asking for, then loads only the single guide that applies. Each guide is written from real usage. It carries exact API shapes, the constraints that are easy to miss, and a "Common Errors" section built from failures developers actually hit, right down to the error text the platform returns. The model recognizes a problem the moment it appears rather than guessing. Take the Job Scheduling skill: point 3 of its prerequisites, exactly as a model reads it, is the difference between a cron job that fires and one that silently never does. That warning exists because someone lost an afternoon to it. A model that has read it never schedules the job that silently never runs. The result is that a model equipped with Agent Skills behaves like a Catalyst solutions expert from the first prompt, with less token consumption, more accurate output, and shorter debugging cycles. The architecture decision A model's default architecture comes from its training data, and training data is full of multi-vendor stacks: one service for hosting, another for the database, and a third for authentication. Those answers aren't wrong; they're averaged. What a model can't know from averages is that on Catalyst, every one of those pieces exists on the same platform, deploys through the same CLI, and shows up on the same bill. So the skills include one file whose job is not code at all. architecture.md is a decision table for the full-stack cloud platform: - Relational data: Data Store, not NoSQL. - A persistent server process: AppSail https://docs.catalyst.zoho.com/en/serverless/help/appsail/introduction/ , not Functions https://docs.catalyst.zoho.com/en/serverless/help/functions/introduction/ . - File uploads: Stratus https://docs.catalyst.zoho.com/en/cloud-scale/help/stratus/introduction/ . Alongside it sit ready-made stack patterns, each with its cost signal attached. The decision table puts the platform's answer in front of the model at the moment the architecture is chosen. The same guide carries the constraints that surface late, like services deprecated in August 2025 that a model trained on older docs will still confidently recommend. For a developer, that is debugging that never happens. For whoever owns the budget, it is one platform bill instead of three vendor subscriptions held together by glue code. This is why we treat the skills as part of building on Catalyst. What that looks like in practice We gave three models the same task: build a full-stack app on Catalyst, end-to-end, once without skills and once with Agent Skills and Catalyst MCP. The strongest model without skills finished 25% of the task. The smallest model with skills finished 92%. The difference was never the model; it was what the model knew about the platform. A note on the numbers: completion counts builds finished end-to-end without a human stepping in, and even with skills, a small share of runs still needed that, which is what keeps the ceiling under 100%. Analysts speak Agentic AI coding agents are good at generating large volumes of code, but assembling the various components of a working application and deploying it into production are still too difficult. With Catalyst, developers can leverage the platform’s Agent Skills, MCP server, and built-in orchestration capabilities to deploy fully functional deterministic applications on the world-class Catalyst PaaS . Any development organization who hasn’t considered Zoho previously should take a close look at Catalyst. — Jason Bloomberg, Managing Director, Intellyx Developers want to work inside the tools they already use. For agents to deliver real business outcomes, they need access to internal systems like CRM and Books, plus the external systems the business depends on, and that data has to stay accurate and current. Catalyst brings compute and logic into one governed familiar environment, giving AI agents the context they need without compromising enterprise control. That turns probabilistic AI-generated code into predictable, maintainable enterprise software. — Shashi Bellamkonda, Principal Research Director, InfoTech Research Group Who this is for - Developers building with Agent Skills ship code that works with Catalyst from the start, with correct handler signatures, correct SDK calls, and correct configuration instead of a plausible draft that needs a debug pass. - Engineering teams and product managers see onboarding time collapse. A new teammate working with a skills-equipped model is productive on Catalyst in their first session. - Students and learners can build on a real, production-grade full-stack cloud platform without first spending weeks in documentation. The model carries the platform knowledge; they bring the idea. - Business and technology leaders get a clearer return on AI investment. An assistant only accelerates a team when it understands the platform the team is building on. Agent Skills make that understanding a given. How to get started Agent Skills work with any AI model and any development tool that supports the open Agent Skills format: Claude Code https://claude.com/plugins/catalyst-by-zoho , Cursor https://cursor.com/marketplace/zoho , GitHub Copilot http://vscode://chat-plugin/install?source=catalystbyzoho/copilot-plugin&plugin=catalyst-by-zoho , Codex https://chatgpt.com/plugins/share/bfcaf08ef64b4f1984849450cfbc2564 , Gemini CLI, and more. Installation is one command: npx skills add catalystbyzoho/agent-skills It detects which AI tools you have configured and installs the skills where each one expects them. From then on, the model has Catalyst's knowledge available in every session. Optionally, connect the Catalyst MCP Server https://docs.catalyst.zoho.com/en/ai-toolkit/catalyst-mcp-server/connect-catalyst-mcp/dynamic-mcp/connect-dynamic-mcp-claude/ . With it, the assistant can go beyond writing code and manage your Catalyst project directly, creating tables, running queries, and configuring resources from the same conversation. The full set of skills is open source at github.com/catalystbyzoho/agent-skills https://github.com/catalystbyzoho/agent-skills . Where this goes next This is the beginning. From here on, every feature we ship to Catalyst carries two documentation tracks: the traditional docs for people reading and learning, and an updated set of skills for models working and building. Every release, every new API, and every architectural change lands in both. Our goal is that a developer using AI on Catalyst is never working with a model that doesn't understand the platform. We are also building this in the open. The format, the depth, and the coverage are actively evolving, and contributions are welcome: edge cases you've hit, patterns worth capturing, or feedback on what's missing. Each one makes the next developer's first prompt better. The bigger picture We are building the layer that sits between Catalyst and every AI model a developer on our platform will ever use, and making sure that layer is always current, always accurate, and always shrinking the distance between an idea and a working implementation. The adoption curve for AI in software development is steep, and it is accelerating. Teams whose models understand their platform will move faster, ship better, and debug less than teams whose models are guessing. That is the gap Agent Skills closes. Cancel https://catalyst.zoho.com/catalyst/blog/catalyst-agent-skills-teaching-ai-agents-how-to-build-on-catalyst.html respond Leave a Reply Your email address will not be published. Required fields are marked