How ConsultingDAG works

The Orchestrator agent reads your asset and plans a directed acyclic graph of specialist agents — analyst first, then executives like CTO, CFO, and CMO chosen for the asset. It returns strict JSON, which armorDag hardens into a valid, laid-out graph (whitelisted agent kinds, deduped ids, exactly one human checkpoint) rendered on a React Flow canvas.

Direction is inferred from your prompt: an existing repo leads with technical due diligence; a fresh idea leads with market analysis. No key? A deterministic heuristic planner builds a comparable DAG so the app is demoable offline.

Both the orchestrator and every specialist run on Gemini through the ADK's LlmAgent + runner. Agents are equipped with FunctionTools — a github_repo reader to ground technical analysis in the real codebase, and a calculator for unit economics. Each node's real token usage is captured and surfaced in the UI.

This is the collaborative core. When a deliverable hinges on a call only a founder can make — pricing model, native vs web, raise vs bootstrap — the agent pauses and asks you in chat, with clickable options (always including an inclusive “you decide” fallback). Your answer is threaded into that agent's work and shared with the rest of the team, so no one re-asks a decision you already made.

A marketing agent can also offer to generate an ad poster as a separate approval — you choose whether to create it.

Guardrails are real ADK callbacks, not a wrapper. Every agent carries beforeModelCallback / afterModelCallback / beforeToolCallback that block prompt-injection attempts, redact secret- and PII-shaped strings from model output, and veto unsafe tool calls (e.g. a github_repo call to a non-GitHub target). A separate structural validator (armorDag) hardens the generated graph itself.

The team shares a genuine ADK BaseMemoryService, backed by SQLite. Each agent writes its finished deliverable to memory (addSessionToMemory) and later agents recall relevant prior work (searchMemory) — persisting across sessions and restarts, the local counterpart of the platform's managed Memory Bank. Completed runs are separately saved to a Blueprint library you can reload from the app.

Every node emits a typed artifact — Markdown, a comparison table, a mermaid diagram, a slide deck, or a generated image — rendered in its detail panel. When the run completes they compile into a single Business Blueprint report (with the collaboration Q&A and a neutral cost estimate) that you can preview and export as Markdown.