Skip to content
Ayan Putatunda
Back to work

Axiomic AI agent harness

open source

Open source, personal R&D, 2025–2026

The multi-agent repository template every Axiomic AI product is built from: agent prompts, MCP server configs, CI, and deploy wiring. Fork, describe the domain, ship.

Claude CodeClaude Agent SDKMCPGitHub ActionsNext.jsPostgresVercel

Problem

Building PersonasFlow taught me that most of the effort in a solo AI product is not the domain, it is the scaffolding: how agents are prompted, which tools they can reach, how work gets reviewed and deployed. Doing that from scratch per product meant the second product would take as long as the first.

Approach

  • Extracted the scaffolding into an MIT-licensed template: an orchestrator plus specialist sub-agents (spec, build, test, review) with their prompts under version control.
  • MCP server configs for the systems every product needs (database, GitHub, deployment) so a fresh fork has working tools on day one.
  • Conventions live as skills the agents load on demand, the same pattern I use on the Zendesk Data Platform.
  • CI runs the agents' own tests before a human reviews the PR; nothing merges without a person.

Impact

  • PersonasFlow, EstateVision, and WealthPilot all run on it; each new product started from the same fork.
  • Axiomic AI operates as an experimental enterprise: the harness is the company's engineering team.

Why this way

A harness is a bet that the interesting part of a product is the domain, not the plumbing. Keeping the plumbing shared and boring is what makes it possible to ship three products alone.

Links