Agent Smith

Inspiration

AI agents are getting easier to demo—but harder to operate.

Every team we talked to could spin up an agent. Almost none had a clean way to deploy it, support it, and evolve it once real users started making requests. Support fell into Slack. Changes were ad hoc. Engineers became the bottleneck.

Agent Smith was inspired by a simple idea: AI agents should ship and be maintained like real software products.

What it does

Agent Smith is a deployment and operations layer for code-first AI agents.

Each deployed agent ships with:

An AI Customer Support Agent that turns user or client requests into structured GitHub issues

An AI Developer Agent that handles bugs and small feature requests using PDD and coding agents like Claude Code or Blackbox

The result: requests flow into real dev workflows, not inboxes—and teams can scale without becoming a support desk.

How we built it

We built Agent Smith around a plan-first, artifact-driven kernel:

A shared agent runtime that separates intent, planning, approval, and execution

GitHub-native workflows for issues, diffs, and reviews

AI agents that write plans and code, while humans retain control over execution

A pack-based architecture so agents can run locally or in the web runtime

Everything is designed to work with real repos, real code, and real teams.

Challenges we ran into

Avoiding the “agent builder” trap while still being approachable

Designing AI agents that help without bypassing engineering discipline

Keeping scope tight while supporting both internal teams and early client deployments

Balancing automation with human approval so trust stays intact

Accomplishments that we're proud of

Built a working end-to-end flow from user request → GitHub issue → AI-assisted fix

Proved that AI support and AI dev agents can operate inside standard software workflows

Designed a generic kernel that supports multiple agent packs without changing core logic

Shipped something opinionated instead of generic

What we learned

Most teams don’t need more agents—they need better process around agents

AI is most effective when it writes plans and code, not when it owns execution

GitHub is still the best system of record for collaborative AI + human work

Constraints make AI systems more trustworthy, not less powerful

What's next for Agent Smith

More agent packs for common internal and client use cases

Deeper GitHub automation with safer approval flows

Observability and trace tooling for agent runs

Making it easier for teams to deploy agents locally or in hosted environments

Our goal is simple: turn AI agents from demos into durable, maintainable products.

Built With

• claude
• pdd