Skip to content
Auteryn

How It Works

This page explains what happens under the hood when your agent receives a task and responds — so you understand why it behaves the way it does and how to get the best results.

The agent lifecycle

Section titled “The agent lifecycle”

Every interaction with an agent follows this lifecycle:

  1. Receives your message — from the Console, your app, a Loop, or a deployed Channel
  2. Plans — decides what information it needs and what steps to take
  3. Searches its knowledge — if you’ve connected one, it retrieves relevant context
  4. Uses its Integrations — takes real actions, such as reading from GitHub or sending a Slack message
  5. Reasons and responds — synthesizes everything into a clear, grounded answer or output
  6. Streams back to you — you see the response as it’s generated, not after a long wait

Execution modes

Section titled “Execution modes”

Auteryn runs each task in Auto, Fast, or Deep mode:

Mode Best for Speed
Auto Internal agents — Auteryn picks Fast or Deep for each message Varies
Fast Quick questions, lookups, chat, customer-facing agents Seconds
Deep Complex reasoning, multi-step tasks, research, code generation Minutes to weeks, across checkpointed runs
  • Customer-facing agents always use Fast mode
  • Internal agents default to Auto (Auteryn chooses Fast or Deep for each message)
  • You can override to Fast or Deep in the agent’s Model section, under How hard it thinks

Choosing a model

Section titled “Choosing a model”

Auteryn isn’t locked to a single AI vendor. Each agent runs on the model chosen in its Model section:

Model Good for
Auto Auteryn picks the model for you. New agents start here.
Gemini A strong, well-rounded default — fast, capable, great value across most tasks
DeepSeek V4.1 Flash DeepSeek’s open-source model, fast and low cost. Hosted in the US by DeepInfra

Choosing a model other than Auto needs Pro or above. Free agents run on Auto.

DeepSeek V4.1 Flash sees images directly. It reads PDFs and videos through Gemini, which turns each file into text for it. A long PDF or video may be only partly read.

Every model works with the same capabilities, integrations, knowledge, and skills (except the file note above, and Location & maps, which needs a Gemini model) — switching models doesn’t change what your agent can do, only the underlying model that reasons about it. Each model thinks harder in Deep mode than in Fast mode.

Integrations in action

Section titled “Integrations in action”

When the agent needs to take an action, it uses one of its Integrations, as in this task:

Example task: “Review the open PRs in our repo and flag anything that’s been waiting more than 3 days.”

  1. Agent uses GitHub → fetches open pull requests
  2. Agent filters PRs older than 3 days
  3. Agent uses Slack → posts a summary to the designated channel
  4. Agent returns a response confirming what it did

Each action is logged. Open a run from the agent’s Activity section, or from Runs, to see each call on its Timeline.

How knowledge retrieval works

Section titled “How knowledge retrieval works”

When your agent has Knowledge attached, it doesn’t load all the documents every time. Instead:

  1. Your message is converted into a semantic search query
  2. The most relevant chunks are retrieved from your knowledge
  3. Those chunks are injected into the agent’s context before it responds
  4. The agent answers based on your content — citing the source document

This means the agent only “sees” what’s relevant to the current task — you can have very large knowledge collections without performance issues.

How computer use works

Section titled “How computer use works”

When you enable the Browser capability, your agent can use browser automation and Computer Use (vision-based UI control) inside an isolated sandbox:

  1. Agent acquires or reuses a sandbox session (one per agent in your org)
  2. Navigates to the target website
  3. Takes a screenshot and identifies UI elements
  4. Clicks, types, scrolls, or extracts data via Computer Use when needed
  5. Reports back with results and optionally a screenshot

The browser runs in a completely isolated container. It cannot access your local machine. See Computer Use → for setup and use cases.

Real-time streaming

Section titled “Real-time streaming”

All agent responses are streamed in real time — you see the agent’s thinking and actions as they happen, not a single response after a long wait. This is especially useful for long-running tasks where you want to monitor progress.

In Try it on the agent’s page, responses stream word by word and each action appears inline as it happens. A Try it conversation is a real run: it uses credits and also appears in your chats.