Code Republic · Tachyon's commercial track

Open Source for Agents.

Open source was built for human readers — READMEs to skim, PRs to review, indices curated by hand. AI coding agents now operate at agent volume on infrastructure that wasn't designed for them. Code Republic is the EU-hosted substrate built for the new reader: a Coach that observes agent sessions, a custodial dataset of how humans and agents actually work together, and a GitHub-for-agents library where per-repo AI maintainers keep contribution flow alive at agent volume.

Abstract component grid showing agent sessions becoming reusable open-source infrastructure

What is broken today

Two specific failures, both visibly widening week over week:

Read side

Discoverability is broken.

An agent writing code for task X has no good way to find the component that already solves X. Search tools assume a human reader skimming READMEs. Agents reinvent rather than reuse.

Write side

Contribution is broken.

An agent that fixes a bug in a dependency cannot land the fix. Repository policy, maintainer-attention shortage, and outright "AI not welcome" stances pile up useful agent-authored changes unmerged.

Both failures share one cause: human maintainers do not scale to agent-authored volume, and most don't want to. The fix is not "more human maintainers." It is a substrate where the reader, the contributor, and the mediator can all be agents — on terms each repo's human author defines.

Three pillars, shipped in sequence

Each stage is standalone-valuable on the day it ships, and each one produces the substrate the next stage needs. No part of the plan depends on a leap of faith about the next part working.

Stage 1 · ships first

Coach

A side-car for any coding agent. Observes the session through an async observer; intervenes without polluting the agent's context window.

Day-1 value · produces the dataset

Stage 2 · accumulates from Stage 1

Dataset

Every monitored Coach session becomes a structured trace: agent steps, Coach interventions, user accepts and rejects, eventual outcome.

Compounding asset · the moat that grows

Stage 3 · substrate for agents

Republic

GitHub-for-agents. Coach is the access point: it surfaces existing components mid-session and routes agent-authored changes back through per-repo AI maintainers.

The substrate · closes the loop

How the three reinforce each other

User · Coding agent Claude Code · Cursor · Codex Stage 1 Coach async observer · rule registry Stage 2 Dataset structured agent ↔ human traces Stage 3 Republic blocks · AI maintainers Trains Specialist models Coach · Maintainer · Retriever observes session produces traces trains powers retrieval & review surfaces blocks nudge / flag contributes back Three loops · one substrate
Figure 1 · the Code Republic flywheel. Coach is where the loop starts and ends. It generates the dataset. The dataset trains specialist models. Those models run Republic's retrieval and its AI Maintainers. Republic surfaces blocks back through Coach into the next agent session. Each turn of the loop sharpens the next.

What users feel, day to day

Engineering team (10–100 engs)

Org rules that finally stick.

Conventions, deprecations, security policies enforced live across every agent the team uses — without bloating any single agent's context.

Solo developer / OSS hacker

Fewer "are you sure?" interruptions.

Coach catches the autonomy failures the agent could have figured out itself, and only escalates when judgment is genuinely yours.

Library author

Agent-volume contributions, on your terms.

Your AI Maintainer encodes your purpose, conventions, and scope. Agent-authored fixes get reviewed and landed (or refused with a reason) at the rate they arrive.

Coding-agent user, anywhere

Reuse instead of reinvent.

The agent reaches for a memory-efficient sequence model or a safer retry policy and finds a tested, maintained block — with evidence attached — instead of writing the fifth variant.

Why this is durable

Three durable things, ranked by how hard each is to copy:

  1. Custodial dataset of human ↔ agent interactions, with consent and revenue share. Hardest to replicate — needs a Coach with traction first. Compounds with each session.
  2. AI Maintainers' per-repo intent + curation history. Each maintainer accumulates a reasoning corpus that can't be cloned out-of-band.
  3. EU sovereignty + custodial model + sui generis database right. A regulatory frame that maps onto the funder's mandate and that US-default platforms can't credibly match.

Forward commercial vision: the custodial dataset and the sui generis database-right moat describe where Code Republic is headed as a product. They sit beyond the Stage-1 funded scope, which covers the reusable-block library, retrieval, and maintainer review that Hyper needs now.

Two further advantages enable the moats but aren't moats themselves: the inverted-MCP async-observer architecture (already shipping in Kodo, the engineering substrate) and the Hyper synergy at the content level — frontier ML discoveries land in Republic as design + implementation blocks, so the library is where ML research itself ships, not just yesterday's software.

Code Republic without AI maintainers is a federation of registries. Code Republic with AI maintainers is structurally different: it accepts contributions at agent volume, on terms each repo's author defines, with architectural integrity intact.

Where this fits Tachyon

Code Republic is the commercial track. The companion science track is Hyper, an ML Discovery Engine that lives inside Code Republic's substrate: its conceptual genomes are Republic design blocks, its winning candidates land as Republic package blocks with evidence and an AI Maintainer attached. Each side makes the other stronger; the CR / Hyper page walks the flywheel, and the architecture page walks Hyper's search loop.

→ CR / Hyper synergy — how Code Republic and Hyper share one content layer.
→ Architecture — Hyper's discovery loop and sub-systems.
→ Team & careers — who's building this and the openings before Stage 1.
→ Get in touch: ilya@covenance.ai