launchroom

rapid agent build · vøiddo

a launch room for the hour before shipping.

paste release evidence, let the agent inspect it, call tools, rank blockers, and return owner-safe next actions instead of a vague green light.

readiness 52
mcpblocked
reporead
evalstrict
statusneeds work

sample payload

release evidence

agent output

blockers and next actions

not a chatbot

every answer carries evidence.

function-calling loop · agentic design

the model decides which tools to call and in what order.

Launchroom uses Gemini 2.5 Flash's native function-calling API. The model is given a set of typed tools and issues functionCall requests across multiple conversation turns. Each result is fed back before the next decision. This is a real multi-step agent — not a single-prompt LLM wrapper. If a partner MCP is unauthenticated, the evaluator reports that honestly instead of hiding it under polished copy.

evidence · not a simulation

real gemini execution trace.

The trace below was produced by running agent/launchroom_agent.py using a real Gemini 2.5 Flash function-calling loop — the model issued 4 tool calls across multiple turns, each executed locally with real API responses. Fivetran partner MCP was live-probed against a real account. This is the exact JSON the agent wrote to disk, not a simulation.

loading trace…