Reply to your message

AgentSquad Architecture Review

You asked how we see this project getting done and which technologies are involved. Here is the full picture.

RemoteFixed scopeNDA firstRead-only accessEnglish or Spanish

AgentSquad is a multi-agent platform running in a private beta. We are looking for an architect who reads the code, works out how it really behaves, and explains it clearly to two audiences: engineers, and people who will never read code.

So the technologies are already chosen. What we want from you is understanding, judgment and clear explanation.

What the system does

Every request follows the same main path. Tracing this path through the code is the heart of the work. Switch the view to see the two ways we need it explained.

  1. Request

    Someone asks for something, by chat or by voice.

    Chat or voice input (Deepgram) reaches the Hono API. The request belongs to exactly one workspace.

  2. Graph

    The system plans the work as a set of steps and the order between them.

    An orchestrator turns the request into a dependency graph of steps, a DAG, from a versioned template.

  3. Steps

    Each step is a known task, such as research, analysis, writing or video.

    Each node runs a versioned operation from a catalog. Run state and an append-only event log live in Postgres; Inngest schedules the workers.

  4. human gate

    Approval

    Some steps stop and wait for a person to approve before the work goes on.

    Gate nodes pause the run until an authorized person decides. The wait is durable: it survives restarts and is recorded.

  5. Deliverable

    A finished piece of work comes back: a brief, a report, a video.

    Outputs are stored per workspace. Row-level security in Postgres keeps each customer's data apart.

Technologies in the code

This is what you would be reading. It is one TypeScript monorepo.

Front end
SvelteKit
API
Hono on Bun
Data
Postgres, pgvector, row-level security
Durable execution
Inngest
Models
LLMs through an AI gateway
Voice
Deepgram
Tools
MCP-based tool layer
Language
TypeScript end to end

You will not add, replace or configure any of these. Use whatever you like to read code and draw diagrams. We only ask that each diagram comes with an editable source file.

How we see it getting done

This is our starting point, not a rule. If you would run it differently, tell us how.

  1. Kickoff call

    We sign the NDA, you get read-only access, and I walk you through the product and the repository layout.

  2. Read and map

    You trace the main path end to end in the code and draft the architecture map: components, data flow, where state lives and where people step in.

  3. Mid-point check

    We review the draft map together, to catch wrong assumptions before you start writing.

  4. Write

    You write the two versions of the document and the honest assessment.

  5. Walkthrough

    You present the architecture in a recorded call, and we close open questions.

What you'd deliver

Architecture map

Done whenOne diagram of components and data flow, and one of the main path, each with its editable source.

Document for engineers

Done whenA new engineer can check every claim against the code, because each one points to the files behind it.

Document for non-technical readers

Done whenA client or investor understands what the platform does and why it can be trusted, without any jargon.

Honest assessment

Done whenIt says plainly what is solid, what is risky and what you would simplify. Blunt is better than polite here.

Recorded walkthrough

Done whenA call of about an hour walks through the map, and we keep the recording.

Access and boundaries

You will get

  • Read-only access to the repository, after the NDA.
  • The existing docs: concepts, architecture notes and runbooks.
  • A live demo of the beta.
  • Direct time with me for questions.

Out of scope

  • Writing features or changing code.
  • Refactoring or infrastructure work.
  • Access to production systems, credentials or customer data.

Next step

If this fits what you do, reply with these three things and we will set up a first call:

  1. A system built on LLMs or agents that you designed or reviewed, and what your role was.
  2. A document or diagram you made that explains a complex system to others.
  3. Your fixed quote for this scope, how many weeks you would need, and when you could start.