Interstice

One command launches your product. AI employees research competitors, create content, send emails, and make real phone calls — then learn from every result to do it better tomorrow.

Comment

Inspiration

Every hackathon has a hundred AI wrappers — a chatbot with a nicer UI. We wanted to build something that actually works the way a company works. Real companies don't have one person doing everything. They have specialists who talk to each other, remember what happened last week, and get better at their jobs over time. We asked: what if a solo founder could spin up an entire launch team with one sentence?

What it does

You type one command — "Launch macroscope.com" — and Interstice does the rest. A CEO agent breaks your command into tasks and delegates to specialist agents: Research, Content, Outreach, and Calls. They all run in parallel, but they're not siloed — the Research Agent publishes competitive findings to a shared channel, and every other agent reads that data before doing their job. Content creates TikTok slideshows using real market data. Outreach writes emails citing actual competitor weaknesses. The Call Agent dials real phone numbers and has real conversations. Every high-stakes action (sending an email, making a call) pauses at an approval gate so you stay in control. And everything that happens gets written to memory — so the next time you run a launch, every agent is smarter than the last time.

How we built it

Each agent runs as a Claude CLI subprocess with persistent sessions — no API keys, no context window resets between tasks. Convex powers our real-time backend: task queue with atomic checkout (no two agents grab the same task), inter-agent message bus, and live dashboard subscriptions with zero polling. Auth0 handles identity at every layer — dashboard login, Machine-to-Machine tokens with scoped permissions per agent, and Token Vault for secure Gmail OAuth. Airbyte syncs lead data from Google Sheets into BigQuery so the Research Agent can query structured data alongside live web results from Perplexity. The Content Agent pushes TikTok drafts to Postiz and tracks what performs to improve next time. Bland AI makes real outbound phone calls with live transcript streaming. The frontend is React with Convex subscriptions — agents light up on the org chart in real-time as they work.

Challenges we faced

  • Inter-agent timing. Content needs Research findings before it can write — but we want parallel execution. We solved this with a pub/sub findings channel: agents subscribe and act when data arrives instead of blocking.
  • Auth0 scoping. Getting M2M tokens with per-agent permission boundaries took careful OAuth scope design. Wiring Token Vault for Gmail through the approval gate system was non-trivial.
  • Airbyte destination issues. Our first destination (Neon Postgres) timed out repeatedly. We pivoted mid-hackathon to BigQuery — then hit IAM propagation delays that cost us an hour before permissions took effect.
  • Memory architecture. Balancing shared company memory, per-agent skill files, and the findings channel without agents overwriting each other required a clear read/write hierarchy.

What we learned

The hard part of multi-agent systems isn't making one agent smart — it's making five agents collaborate. A single brilliant agent is less useful than five agents that share context and build on each other's work. The compounding memory system ended up being our strongest feature: agents that accumulate knowledge create a flywheel that single-prompt tools can never match.

Built With

Share this project:

Updates