AI-native
product lab.

Configuration comes down to four decisions: which signals it reads, which tools it can touch, the voice it writes in, and where the permission line sits. We scope a private retrieval index over docs, Slack, and email; wire permissioned tool access (Linear, Notion, Stripe, GitHub) through MCP servers; tune voice and tone to the founder; and run a weekly eval-and-retrain loop so it stays calibrated. It ships on the Claude Agent SDK against current Claude models, with provider fallbacks underneath. The agent drafts; the founder signs.

Treat an MCP server like any other production surface. OAuth and least-privilege scopes per tool, secrets that never reach the model's context, rate limits and idempotency on anything that writes, versioning so a tool change doesn't silently break an agent mid-session, and traces on every call. The Model Context Protocol makes capabilities portable across agents — which is exactly why a sloppy server becomes everyone's incident.

Route by job, not by habit: a fast model for classification and extraction, a frontier model — Claude Opus 4.8 today — for hard reasoning, cheaper tiers for bulk work. Put provider fallbacks underneath so a refusal, a rate limit, or a 529 reroutes inside the same request instead of failing the user. Add cost and latency budgets, prompt caching on the stable prefix, and graceful degradation when the best option is unavailable. The user should never know which model answered.

Write the golden set before the prompt: representative inputs, expected outputs, and a pass/fail rubric per case. Run every prompt and model change against it, gate the deploy on the result, and feed real production traces back in so the set grows toward the cases that actually break. When the model regresses you get the commit that did it; when it improves you get proof instead of a vibe.

Structure traces around the loop: every prompt, every tool call, every model response, with token and cost attribution per step and per tenant. Alert on what actually hurts — runaway loops, climbing latency, failure spikes, cost drift — and redact PII at the boundary so the logs are safe to keep. Pair it with a token-and-cost dashboard an operator can read without a data team. Observability is what turns the vague 'the agent feels slow' into a fix.

The whole build, end to end: scaffold the Claude Agent SDK, wire your calendar, inbox, and Slack through MCP servers, keep it read-only with a tool whitelist, write the brief prompt, and schedule a 7am run. Real commands, grounded in the current SDK and MCP docs. The last twenty percent — voice, retrieval, evals — is what turns a demo into something a founder trusts.