Building a Multi-Model AI Platform and Agent Marketplace

quora · Tech · 2025 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Consumer multi-model AI platform and bot marketplace

Quora needed a new AI product paradigm after GPT-3 experiments showed that chat, not Quora's original library-style interface, was the right way for many users to access model-generated knowledge, diverse models, custom bots, and AI applications.

Constraints:

• Quora was not building the foundation models itself
• users and developers wanted access to many models, modalities, and agents through one interface
• model providers can go down and users do not necessarily accept transparent substitution
• bot creators need client distribution, monetization, authentication, billing, and scale without building all of it themselves
• the AI ecosystem changes too quickly for multi-year planning horizons

Architecture

Poe acts as a common consumer interface for many AI models and bots. It provides native clients and web access, a single subscription/points system, prompt bots that wrap a base model and prompt, server bots that call creator-hosted HTTP endpoints on each user message, parallel multi-model querying through @ mentions, bot discovery and sharing, monetization, and distribution to millions of consumers. The platform currently emphasizes read-only artifacts and plans future support for side-effecting agents.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  poe-client["Poe client"] --> poe-platform["Poe platform"]
  poe-client --> raw-models["Raw models"]
  poe-client --> prompt-bots["Prompt bots"]
  poe-client --> server-bots["Server bots"]
  poe-platform --> raw-models
  poe-platform --> prompt-bots
  poe-platform --> server-bots
  poe-platform --> points-and-monetization["Points and monetization"]
  poe-platform --> bot-marketplace["Bot marketplace"]
  poe-platform --> app-creator["App creator"]
  parallel-model-query["Parallel model query"] --> raw-models
  parallel-model-query --> prompt-bots
  prompt-bots --> raw-models
  prompt-bots --> bot-marketplace
  server-bots --> creator-servers["Creator servers"]
  server-bots --> bot-marketplace
  bot-marketplace --> points-and-monetization
  app-creator --> prompt-bots
  app-creator --> server-bots

Components

Data flow

A user opens Poe through a web or native client and either chats with a raw model, invokes multiple models in parallel with @ mentions, uses a prompt bot that combines a base model with stored instructions, or sends a message to a server bot. For server bots, Poe sends an HTTP request to the creator's server and inserts the returned result into the Poe message stream. Subscription points and creator-set pricing track usage and revenue sharing, while the creator can focus on the model, workflow, or application logic.

Why this architecture

Poe was shaped as a browser-like common interface because Quora expected a diverse ecosystem of models and AI products, and developers should not have to build every client, protocol, monetization system, and consumer distribution channel themselves. Prompt bots support low-friction customization, server bots support complex workflows and hosted models, and parallel querying serves users who actively compare model behavior.

Models used: GPT-3 · GPT-4 image generation · GPT Image 1 · OpenAI o3 · OpenAI o4 mini · Gemini 2.5 · Claude Sonnet 3.7 · DeepSeek models · Retro Diffusion · text models · image models · video models · audio models

Scale

• Poe provides iOS, Android, Windows, Mac, and web clients.
  “can just plug into PO and we provide uh iOS, Android, Windows, Mac clients, a web interface. We provide monetization. Um, we provide history and sort of all this work that you need to to get a an AI product to millions of consumers around the world. Um, we just sort of”transcript, stated
• Poe aims to get AI products to millions of consumers around the world.
  “Um, we provide history and sort of all this work that you need to to get a an AI product to millions of consumers around the world. Um, we just sort of have a a single place to to do all of that. Yeah. And I I think one of the interesting things that we'll probably”transcript, stated
• Creators are making millions of dollars per year on Poe.
  “there's there's there's companies making in the millions of dollars per year through the millions of dollars per year.”transcript, stated
• Individual creators can make hundreds of thousands or tens of thousands of dollars per year.
  “there's like individual people making in the in the hundreds of thousands and tens of thousands of of dollars per year.”transcript, stated
• Quora's planning horizon for Poe is about two months because the AI ecosystem changes so quickly.
  “We I'd say our plans go out about two months right now. Okay, that's that's how that's how far we have. We have some very hopefully you you'll all see some some great stuff launch in the next few months. Um and then we will”transcript, stated
• Quora started in 2009 and launched its first version in 2010, contrasting with the faster Poe planning environment.
  “uh we started in 2009 and launched the first version in 2010 10. It was I'd say like it took maybe like every few years there was a significant shift. You know, there was a big shift with mobile and there was a sort of shift with with”transcript, stated

Alternatives considered (roads not taken)

Integrate GPT-generated answers into Quora's existing interface

Rejected because: Quora's experiments suggested chat was the better paradigm for humans getting knowledge from LLMs.

Build a single-model product

Rejected because: Quora bet on a wide diversity of models and products, and Poe's users often specifically want multiple models.

Expect users to switch transparently when a model provider is down

Rejected because: Users are attached to specific models and become unhappy when the model they rely on is unavailable.

Only support prompt bots

Rejected because: Complex agents, outside data sources, custom inference, and hosted models require server bots.

Support side-effecting agents immediately

Rejected because: The current platform mostly supports read-only artifact-producing bots while acknowledging real-world actions carry risks and require more platform work.

Patterns

model-routing (routing)

Poe exposes many foundation and specialized models under one interface and lets users explicitly route a prompt to one or many models.

tool-use-orchestration (agents)

Server bots let developers connect Poe conversations to external servers, data sources, hosted models, and more complex agent applications.

cost-aware-routing (operations)

Poe's creator economy converts creator-set dollar prices into points and passes usage costs through to users based on subscription spend.

prompt-versioning (operations)

Prompt bots make a prompt reusable and shareable instead of forcing users to repeatedly paste the same prompt into a base model.

Tradeoffs

Product interface

Chose build a new chat-first product over fold GPT answers into Quora's existing Q&A interface.

Rationale: Quora concluded LLM knowledge access was a paradigm shift toward chat-style interaction.

Platform positioning

Chose common interface for diverse models and AI products over owning one foundation model.

Rationale: Quora was not building LLMs and believed the ecosystem would contain many models and products needing distribution.

Bot creation tiers

Chose prompt bots plus server bots over one creation path for all bot builders.

Rationale: Nontechnical prompt creators and sophisticated developers/model hosts need different degrees of control and infrastructure.

Agent capability scope

Chose read-only/artifact-producing bots for now over side-effecting real-world actions immediately.

Rationale: Poe wants to support side effects eventually, but current priorities and risks keep most bots read-only.

Planning cadence

Chose short two-month planning cycles over multi-year platform roadmaps.

Rationale: The AI model, modality, agent, and tooling environment changes too fast for stable long-range plans.

Gotchas (what bit them)

Users do not treat provider downtime as interchangeable across models.

Root cause: Power users are model-specific and often depend on a particular model's behavior.

Resolution: Poe exposes multiple models and comparison workflows, but does not assume users will accept silent replacement when a provider is down.

Prompt bots are simple to create but not simple to make good.

Root cause: Effective prompting requires empathy with model behavior and persistent testing across many cases.

Resolution: Poe supports prompt bots as a low-friction tier while server bots exist for more technical creators and complex applications.

Text models dominate despite support for image, video, and audio.

Root cause: Text model quality and economic value are currently higher than image/video quality for many practical tasks.

Resolution: Poe keeps a general interface and follows the market as modality quality changes.

Long-term planning is hard in the AI platform environment.

Root cause: New models, modalities, agents, MCP, and tools appear every month.

Resolution: Poe plans about two months out and continually adapts after launches.

Side-effecting agents need more safety and product work than read-only bots.

Root cause: Actions that modify external systems carry obvious platform risks compared with artifact generation.

Resolution: Poe currently focuses mostly on read-only bots and treats MCP/standards as part of the future path toward side effects.

Evaluation

Poe uses product telemetry and direct user behavior as evaluation signals: users compare multiple models in parallel, surveys identify why people use Poe, provider downtime reveals model-specific dependence, and usage trends show which modalities and reasoning models are gaining traction.

Metrics: model usage trends · parallel model comparison behavior · survey responses about why users use Poe · creator revenue · subscription points consumed · modality usage mix · provider downtime sensitivity

Outcomes

• Poe provides one interface and subscription for multiple AI products and models.
• Creators can build prompt bots, server bots, hosted-model bots, and more complex agent applications.
• The marketplace has companies making millions of dollars per year and individuals making hundreds of thousands or tens of thousands per year.
• Users can compare multiple models in parallel and use specialized models and bots through one consumer distribution layer.

Poe distributes AI products through web and native clients to millions of consumers, supports a marketplace economy with creator pricing and points, and plans on a roughly two-month horizon because new models, modalities, MCP, agents, and tools arrive every month.