Work · Tarva · 2025–2026 · Operational · daily use
Tarva’s own infrastructure · supports product delivery · not a client engagement
AgentGen
140+ specialist agents, reproducible and versioned.
Availability
AgentGen is operational and in daily use inside Tarva, so it has no public link.
The problem
Generic prompts produce generic work, and a delivery system that needs a UX architect, a patent strategist and a security engineer on each project needs those experts to be dependable and easy to find.
The system
Tarva's agent factory and registry: a Python CLI that generates specialist agents, a runtime registry of agent definitions with methods, skills, templates and quality rubrics, semantic search, signing and benchmarking.
Outcome
Each project can call on reproducible, versioned specialist agents instead of improvising an expert again in every session.
Proof
AgentGen is operational and used daily by three companies, including inside Serras.
Role
AgentGen is Tarva's own system. Tarva led it with AI assistance across agent definitions, skills, semantic routing, signing, modernization, benchmarks and lifecycle.
Supporting stack
Python CLI · Markdown/YAML · Supabase pgvector · Ollama · TypeScript MCP
Scale
140+ agent definitions · 3 companies daily