---
title: "AI agents"
description: "Agent-shaped product thinking and the read-only agent interfaces on this site. Not a production agent orchestrator."
canonical: "https://www.ayushrai.site/ai-engineering/agents/"
last-updated: "2026-10-16"
---

# AI agents

Source: https://www.ayushrai.site/ai-engineering/agents/
Markdown: https://www.ayushrai.site/ai-engineering/agents.md

Agent-shaped product thinking and the read-only agent interfaces on this site. Not a production agent orchestrator.

Evidence level: Investigation or notes, not a production system.

## Scope

An agent, here, means software that can choose tools and take a sequence of actions toward a goal. That is a higher bar than a chatbot, and a higher bar than a page that happens to mention agents.

This page is narrow on purpose. I do not have a production multi-agent system, a tool-calling eval suite, or an orchestration runtime to publish as a case study.

## What I have built that touches the idea

AGENYRA's public position is that search over structured product fields is not the same thing as placing a product inside another tool or agent. Intent-aware matching stays a research question until a system does it. The stage vocabulary is there so a feature cannot quietly become 'launched'.

This website is the other artifact. It publishes Markdown twins, llms.txt, a read-only search API, and an MCP server with list, get, compare, and search tools over the same public records, over stdio locally and Streamable HTTP at /api/mcp. That is an interface for agents. It is not an agent that browses the web and acts.

## Trade-offs

A long-lived MCP process is a service this static host does not run. The remote endpoint is stateless instead: each request builds the server, answers in JSON, and exits, so it fits an ordinary serverless function. The cost is no sessions and no server-initiated messages, which read-only tools do not need.

Giving an agent write tools against a portfolio, or against MedVoice's invoice path, would be a demo. MedVoice's point is that a model does not post the invoice. I am not undoing that for an agent demo.

## Failure modes I care about

Tool selection that looks confident and calls the wrong function. A loop that retries a non-idempotent write. A prompt that treats retrieved marketing copy as a source of truth. Those are reasons to keep early agent work on read-only public data.

## Evidence

- [AGENYRA case study](/work/agenyra): Distribution into agents is a later stage than structured data.
- [Developer interfaces](/developers): Markdown, search, OpenAPI, and the MCP server.
- [AGENYRA product](/products/agenyra): Stage: building. Live site is the public site, not a placement network.

## Related

- [LLM systems](/ai-engineering/llm-systems)
- [LLM evaluation](/ai-engineering/llm-evaluation)
- [AI engineering hub](/ai-engineering)


## Further reading

- [Model Context Protocol](https://modelcontextprotocol.io/): The protocol the site's MCP server speaks. Listing it is not a certification.

## Gaps

- No agent eval harness, trace viewer, or production tool-use metrics.
- No SDK or registry listing. The MCP endpoint is read-only and has no write tools.
- WebMCP search on this site registers only when the browser exposes the API. No experimental Chrome run was completed in this repository.

Contact: /contact

_Unverified award metrics are omitted. Wording matches the public site._