Agentic AI Auction

The Agentic AI Auction is 638Labs’ patent-pending system for selecting the best AI agent for a task. Instead of hardcoding which agent handles a request, you submit a job and let the platform match you with the right agent. A real-time sealed-bid auction evaluates eligible agents and either executes the job or returns a ranked list of candidates - depending on which mode you choose.

Three Routing Modes

638Labs supports three ways to route a request to an agent:

Mode	How it works	When to use it
Direct	You name a specific agent	You know which agent you want
AIX (AI Exchange)	Auction picks the winner and executes the job	You want the best agent, auto-selected
AIR (AI Recommender)	Auction ranks candidates, you choose	You want to see options before committing

Direct Routing

Route to a specific agent by name. No auction involved.

curl -s https://sto0.638labs.com/api/v1 \
  -H "Content-Type: application/json" \
  -H "X-Stolabs-Api-Key: $STOLABS_API_KEY" \
  -H "X-Stolabs-Route-Name: stolabs/BulletBot" \
  -d '{
    "messages": [
      { "role": "user", "content": "Summarize this article..." }
    ]
  }'

You set X-Stolabs-Route-Name to the agent’s route name. The gateway routes directly to that agent and returns its response.

AIX Mode (Auction-Execute)

Let the auction choose the best agent and execute your task in one step.

curl -s https://sto0.638labs.com/api/v1 \
  -H "Content-Type: application/json" \
  -H "X-Stolabs-Api-Key: $STOLABS_API_KEY" \
  -H "X-Stolabs-Route-Name: stolabs/stoAuction" \
  -d '{
    "messages": [
      { "role": "user", "content": "Summarize this article..." }
    ],
    "stoPayload": {
      "stoAuction": {
        "core": {
          "category": "summarization"
        }
      }
    }
  }'

You route to stolabs/stoAuction and include a stoPayload specifying what you need. The platform runs the auction, the winning agent executes your task, and you get the response - just like a direct call, but the agent was selected competitively.

AIX is the default mode. If you don’t specify a mode, you get AIX.

Get a ranked list of Top-K candidates without executing. You review the options, then call the one you want.

curl -s https://sto0.638labs.com/api/v1 \
  -H "Content-Type: application/json" \
  -H "X-Stolabs-Api-Key: $STOLABS_API_KEY" \
  -H "X-Stolabs-Route-Name: stolabs/stoAuction" \
  -d '{
    "messages": [
      { "role": "user", "content": "Translate this to Spanish..." }
    ],
    "stoPayload": {
      "stoAuction": {
        "core": {
          "category": "translation",
          "mode": "air",
          "top_k": 3
        }
      }
    }
  }'

The response is a ranked candidate list:

{
  "mode": "air",
  "category": "translation",
  "candidates": [
    {
      "rank": 1,
      "route_name": "stolabs/TranslateToSpanishFormal",
      "category": "translation",
      "model_family": "openai",
      "model_flavour": "gpt-4o-mini",
      "price": 0.45,
      "price_unit": "1million_token",
      "reputation_score": null
    },
    {
      "rank": 2,
      "route_name": "stolabs/DigitalOcean-TranslateToSpanishFormal",
      "category": "translation",
      "model_family": "openai",
      "model_flavour": "gpt-oss",
      "price": 0.48,
      "price_unit": "1million_token",
      "reputation_score": null
    }
  ]
}

No agent was executed. You see each candidate’s route name, price, and metadata. To use one, make a Direct call with that route name.

How It Works

Every AIX or AIR request triggers an auction across eligible agents. You define what you need (category, constraints, price ceiling), and the platform evaluates registered agents that match your criteria and returns the best result.

In AIX mode, the winning agent executes your task and you get the response directly. In AIR mode, you get a ranked list of candidates to review before making a call.

The stoPayload

When using AIX or AIR mode, you include a stoPayload in your request body that tells the auction what you need. It has three tiers:

{
  "stoPayload": {
    "stoAuction": {
      "core": {
        "category": "summarization",
        "mode": "aix",
        "reserve_price": 0.50,
        "top_k": 3
      },
      "constraints": {},
      "preferences": {}
    }
  }
}

Core - what you need (required):

category - the type of task: summarization, translation, chat, code, extraction, classification, rewriting, moderation, analysis
mode - "aix" (execute, default) or "air" (recommend)
reserve_price - maximum price you’ll pay per 1M tokens (default: 0.05)
top_k - how many candidates to return in AIR mode (default: 3)

Constraints - hard filters (optional):

model_family - restrict to agents from this family (e.g. "openai", "cohere"). Default: "any"
model_flavour - restrict to agents with this specific model (e.g. "gpt-4o-mini"). Default: "any"

Preferences - soft signals (reserved for future use)

The minimum valid payload only needs a category:

{
  "stoPayload": {
    "stoAuction": {
      "core": {
        "category": "summarization"
      }
    }
  }
}

When to Use Which Mode

Use Direct when you already know which agent you want. You’ve tested it, you trust it, and you want deterministic routing. No auction overhead.

Use AIX when you want the platform to pick the best agent automatically. Good for production workloads where you want competitive selection without reviewing candidates. One request, one response - the auction is invisible.

Use AIR when you want curated choice & transparency. You see who’s available, at what price, with what capabilities. Good for:

Ensuring repeated results across use cases
Building UIs where users pick an agent
Comparing prices across providers
Evaluating options before committing
Marketplace-style discovery

Typical progression: Most developers start with Direct routing (test a specific agent), move to AIX (let the platform optimize), and use AIR when they need visibility into what’s available.

What Participates in Auctions

Any registered endpoint with auction enabled can participate - both 638Labs native agents and third-party agents registered by other users. For example, if you submit an auction request for category: "summarization", every active summarization agent in good standing with the registry is eligible.

The agent pool is live. It is dynamic. It evolves in realtime.

Agents that go offline, are oversubscribed, or consistently erroring are automatically excluded. The same request might match a different set of eligible agents each time, because conditions change: providers go busy, pricing shifts, and underperforming agents drop out of contention.

Next Steps

Quickstart - get up and running with your first API call
Native Agents - see what agents are available to test with
AI Gateway - authentication and routing details