An inbox that answers back

01 · THE PROBLEM

Small businesses drown in messages, not work

We build websites and run marketing for small businesses. The pattern repeats at every client: the work is fine, the communication is the bottleneck. A lead texts the office line at 9 PM. An invoice lands in a shared mailbox nobody watches. A client asks the same question three people already answered. Nothing is lost, exactly — it is just spread across five apps and four people's memories.

CommSync is our answer: one org-wide inbox where SMS, email, calls, and voicemail land as threads the whole team can see. For this hackathon we took its agentic core, deployed the full stack on Google Cloud, and put Gemini on Vertex AI behind every feature that thinks.

The thesis of the Taskmaster track is an agent that handles complex workflows and takes action. An inbox is the harshest place to test that thesis. The input is unbounded, adversarial, and written by strangers. The action — sending a message to a real customer — is public and irreversible. So most of what we are proud of here is not the model call. It is everything wrapped around it.

02 · THE HEADLINE ACT

Sage: an employee made of function calls

Our demo workspace is Wrenhouse, a fictional design-build studio. Wrenhouse runs an agent named Sage on its office phone line. When a brand-new lead texts in — square footage, a budget, a timeline — Sage reads the message, searches the studio's knowledge base, and answers with a real ballpark and a qualifying question. No made-up prices. No rule was written. That is a native Gemini function-calling loop, running live against tools like search_knowledge, read_thread, and send_reply.

Each agent is a first-class org member of kind AGENT — no login, no seat, but real channel permissions, so the whole access-control and fan-out machinery works on it unmodified. Every turn is inspectable after the fact: what triggered it, which tools it called, what it said, and what it cost, down to the fraction of a cent.

Autonomy is a dial, not a switch. In DRAFT mode every reply waits in an approval queue. SUPERVISED sends but keeps humans in the loop. AUTO is hands-off. And a human can always take the thread back: one click pauses the agent on that conversation, because the fastest way to lose a customer's trust is a machine that will not yield.

04 · AGENTIC RETRIEVAL

Ask: memory for the whole inbox

Ask is a streaming, agentic chat over everything the org has ever said. "Did anyone push back on cost recently?" is not a keyword query — so Ask plans one: it runs semantic search over a pgvector index, reads the candidate threads with its own tools, and answers with citations that deep-link to the actual messages. It reads attachments too. Ask "how much was that invoice?" and the number comes from inside the PDF.

Under the hood it is a tool loop with a per-turn token budget and a hard step cap, over a dozen permission-scoped tools — every tool call is filtered by the caller's real channel access, so an agent can never read a thread its human could not. The same loop doubles as the in-app help desk: it searches the live product docs, so "why did my email bounce?" gets an answer grounded in both the thread and the manual.

05 · THE AMBIENT LAYER

The agents you never see

The loudest demo is the agent that texts a customer. The biggest daily win is quieter: background workers that run on every message with no one watching.

A classifier reads each inbound sender and files machine mail — receipts, alerts, promotions — out of the way, with a deterministic header-analysis layer that catches automated campaigns no matter how personal the prose sounds. Twice a day, a scheduled agent writes The Daily Brief: the day's threads, ranked like a front page, each item grounded in a real message and delivered as a designed email. Thread titles, reply drafts briefed on the whole conversation, even the composer's tab-to-complete typeahead — each is its own small model pipeline with its own budget, dedup hash, and audit row.

Every one of these pipelines shares a harness: content-hash dedup so retries are free, per-org credit accounting on every call, and a kill switch per feature. Agentic systems fail weird; the harness makes them fail boring.

06 · LESSONS

What building this taught us

Fail closed, everywhere, first

The gate rail was not a compliance afterthought — it was the design. In an agentic system the cheapest failure is the model call you never make. Deterministic checks run before inference; inference runs before action; action is idempotent. In that order, always.

The boring rails are the product

Sends are non-idempotent, so three independent layers guarantee at-most-once delivery. Job queues debounce per thread so a burst of messages becomes one thoughtful turn, not five. None of this demos well. All of it is why you can let a machine text your customers.

Agents are users now

CommSync ships an MCP server with 70+ tools, signed outbound webhooks, and machine-readable docs endpoints — external agents are a first-class client of the product, with the same permission model as humans. This page practices what we preach: semantic HTML, JSON-LD facts, and a plain-markdown mirror, because we assume at least one reader of this post is an agent.

08 · FOR THE RECORD

Fact sheet

The judges asked for an agent that takes action. Ours texts your customers back — politely, inside quiet hours, at most a few times a day, and never after they say stop. We think that restraint is the interesting part.

— The CommSync team at Haven Media Solutions