Built to test whether AI can run the day-to-day of a venture studio.
Orbit is the multi-agent platform we built to operate Averian Labs. Nine specialized AI agents across the venture lifecycle. Two founders. Three SaaS products. One public hypothesis being validated toward $500K ARR.
The hypothesis
We started Averian Labs to test a specific question: can two founders, equipped with a multi-agent platform that handles day-to-day operations across the venture lifecycle, run a multi-product venture studio to $500K ARR — replacing work that traditionally required a team of 10 to 20 people?
Orbit is the experimental apparatus. Averian Labs's three products (BuildCreditAI, Laudvi, TradeReply) are the experimental conditions. The $500K ARR target is the validation threshold. Hitting it = hypothesis validated. Falling short = data we learn from and share.
This isn't a thought experiment. It's a real platform running real products. The result is unknown.
The team
Orbit operates as a team of nine specialized AI agents organized in two layers — specialists who do the lifecycle work, and meta-agents who build and coordinate the platform itself.
Specialists
- Olivia — Idea Analyst. Finds and validates opportunities.
- Drew — Product Architect. Turns validated ideas into product outlines.
- Mason — Build Planner. Sequences technical work into sprints.
- Cody — Code Executor. Turns build-plan sprints into Claude Code handoffs.
- Quinn — QA Reviewer. Reviews pull requests, surfaces concerns and recommended focus.
- Lana — Launch Strategist. Pricing, positioning, first-10-sales, launch copy.
- Indigo — Brand and Design Director. Produces brand systems (strategy + visual identity + design tokens).
Meta-agents
- Forge — Builds new agents from founder briefs. Produces specs that Cody executes.
- Helios — Orchestrates the specialist team. Surfaces what needs founder attention.
Day-to-day work being delegated
Each agent represents a category of day-to-day venture studio work being delegated to AI:
- 1. Find
- Olivia (live)
- 2. Validate
- Olivia (live)
- 3. Architect
- Drew (live)
- 4. Plan builds
- Mason (live)
- 5. Build
- Cody (live, code); Reed (not yet built, content)
- 6. QA
- Quinn (live)
- 7. Launch
- Lana (live)
- 8. Brand
- Indigo (live)
- 9. Refine PMF
- Paige (not yet built)
- 10. Improve
- Pending agent
- 11. Learn
- Pending agent
- 12. Scale
- Pending agent
- 13. Repeat
- Loops back to step 1
Currently 8 of 13 steps have a live agent. The remaining 5 (Reed for content, Paige for PMF refinement, the Improve agent, the Learn agent, the Scale agent) are pending — Forge will build most of them.
Forge built Helios
The sharpest current data point: the recursive autonomy moment.
Indigo, the brand director, was the last manually-built agent (June 15, 2026). Forge — the agent that builds agents — was the bridge: built manually, but designed to design. Helios — the team orchestrator — was Forge's first creation. Heather wrote a brief. Forge produced a complete spec for Helios. Heather reviewed and refined it via revision. Claude Code built Helios from Forge's spec. Helios shipped (June 16, 2026).
Even the work of extending the platform is becoming automated. If the hypothesis is “day-to-day can be automated,” recursive autonomy says the meta-work of extending automation is being automated too.
What founders still do
The experimental result so far: founders still spend meaningful time on strategy, judgment, and approval. The promise was never “AI replaces founders.” It's “AI replaces day-to-day execution so founders can focus on irreducible founder work.”
Things agents handle
- Drafting specs (partially — Cowork-Claude assists)
- Code execution, commits, pull requests
- Code review
- Brand systems
- Launch strategy
- Cross-product orchestration
- Idea validation
Things founders still do
- Strategic decisions (which product, which positioning, which audience)
- Approval at lifecycle gates (every agent invocation is founder-gated)
- Pricing decisions (BuildCreditAI's free tier question is currently founder territory)
- Positioning calls (rebranding, naming, voice)
- Customer empathy + product-market fit judgment
- Trust commitments (which brand promises, which partnerships)
The patterns we use
Eight institutional patterns Orbit has codified through real shipping. We share these because they're more valuable to other builders than the source code would be.
- Schema discovery first. Every slice spec queries information_schema before writing migrations. Never invent column names.
- Append-only history. Agent outputs get new rows, never UPDATE. You can audit “what did Olivia say last Tuesday.”
- Strict:false + Zod. Anthropic strict mode rejects integer bounds; Zod is the actual validation gate.
- ASCII-only commits. No smart quotes, em-dashes, ellipses. PowerShell + git compatibility.
- Founder-gated invocation. Helios recommends, doesn't auto-trigger. Cody requires founder-approved handoff. Multiple checkpoints prevent accidents.
- RLS three-policy pattern. SELECT via current_site_ids(), INSERT/UPDATE via has_site_role(site_id, editor/admin arrays). Consistent across every site-scoped table.
- Service-role client for runner/admin paths only. Not the SSR cookie client (which silently forwards user JWT and breaks service-role intent).
- One-site-per-product convention. Each product is a site_id in Orbit's DB. RLS scopes everything. Site selector switches context.
Watch the experiment
Orbit isn't for sale. The platform's value to other builders is the methodology, not the source. We share what we learn through the build-in-public blog (coming soon). If you're a builder applying similar patterns, ask anything in the comments — we read and reply.