Skip to main content
Beta

Flow is in beta. We are actively shaping the product, so things may change as we iterate. Your feedback is welcome.

Agentic AI

The assistant behind Flow is not a single prompt: it is an agent, with explicit instructions, a set of callable tools, and the ability to chain its own actions. Agents, skills, prompts, specs, and playbooks are authored in AI Foundry and surfaced inside Flow through the Catalog page, where you pick which ones to use in your conversations. This page explains the building blocks and how they fit together.

The building blocks

Every building block listed below is a AI Foundry catalog resource. Inside Flow, the Catalog page is the place where you browse them and decide which ones to apply to your conversations:

TypeWhat it isAI Foundry reference
AgentsA model + a tool selection + system instructions. The unit that actually answers your messages.Agent
ModelsLLM models available to agents.Model
PromptsReusable system prompts that shape the agent's tone, role, or workflow.Prompt
SkillsReusable, named operations an agent can perform. Distinct from the tools exposed by Connectors.Skill
Spec TemplatesSpecKit command templates: structured commands the agent can apply to a project.Spec Templates
AI PlaybooksCurated bundles of all of the above, applied as a single unit.Playbook

Agents

An agent has several components:

  • a model with advanced configurations (e.g. temperature);
  • a set of instructions;
  • one or more skills;
  • one or more tools.

Skills are reusable, named operations that agents can perform. They are defined in AI Foundry and are distinct from the tools that come from Connectors (which expose external systems like GitHub or Jira).

Agents are defined in AI Foundry; from Flow you pick which agent answers your conversation.

Different agents are useful for different jobs: for example, a fast agent for quick edits and a slower reasoning agent for complex planning.

AI Playbooks

A Playbook is a named bundle of every artifact above and is the unit you usually apply to a conversation. Playbooks are authored in AI Foundry and made available inside Flow through the Catalog:

AI Playbook
├── Agents (model + tools + instructions)
├── Skills (reusable agent operations)
├── Specs (SpecKit templates)
└── Prompts (system instructions)

A Spec Template lets agents run named commands against a project. When you invoke a command in chat, Flow looks it up in the Catalog and runs the matching spec against the active project.

When a playbook is activated:

  1. All tools from referenced agents are merged into the active tool selection.
  2. System instructions are rebuilt from the playbook's prompts.
  3. The playbook is pinned to the conversation and re-applied automatically the next time you reopen it.

Default playbooks

Under Settings → Advanced, you can pick a Default Chat Playbook and a Default Code Playbook. The chosen playbook is applied automatically when a new conversation starts, so you always open Flow with the right setup for your use case. This is a personal preference and only affects your account.

Reasoning and limits

For heavy reasoning models, Flow hides internal chain-of-thought from the chat by default, so responses stay focused on what is useful to the user. The number of consecutive tool calls per agent turn is capped to keep conversations responsive: if a task needs more steps, the agent splits the work across multiple turns.

Putting it together

A typical workflow looks like this:

  1. In AI Foundry, define Skills for the operations agents should perform, group them into Agents, wrap recurring tone or workflow choices in Prompts, and add Specs for repeatable commands.
  2. Bundle everything into a Playbook.
  3. Open Flow, pick the playbook from the Catalog, and start a conversation: the assistant is already configured for the use case the playbook was built for.

If you want a playbook to be applied automatically every time you open Flow, set it as your default in Settings → Advanced.

See also

  • Chat: where playbooks are applied to a conversation.
  • Connected tools: the external systems agents can act on.
  • Code: the Canvas where agent output becomes a running project.