Why an agent-native portfolio?

Date: 2026-04-18 Post 1 (build-in-public)

Recruiters and hiring managers at AI companies are starting to run the same workflow as their engineers: connect an agent, let it research, skim the highlights. If the portfolio is a static site, the agent has to screen-scrape markdown and pray it parsed right.

So I'm building a portfolio that IS an MCP server. Connect your agent, call structured tools, get structured answers. The same handlers also power a CLI and a web UI. Four interfaces, one handler.

The portfolio is the product demo. If the thesis of "harness engineer" means anything, it's that the scaffolding around the model matters — the tool choices, the data shape, the guardrails, the evals. Shipping a portfolio without those isn't demonstrating harness engineering, it's describing it.

This week

• Phase A (today/tomorrow): Turborepo monorepo, pre-commit sanitization gate, workspace scaffolding. Boring but load-bearing.
• Phase B (days 3-6): Real MCP server with 9 tools, the CLI, first HTTP endpoint.
• Phase C (days 6-7): Web scaffolding. Landing page shell.

Where to look

The agent at the top of this page is the live version. The eval harness behind it is open source: agent-evals.

— Daniel