Agents

Agents in Dezifi

An Agent is your unit of intelligence — one LLM, the Tools it can call, the Skills and knowledge it draws on, and the Guardrails that constrain it. Build one with a 9-step wizard, test it in chat, ship it as an API or a Slack bot.

What you'll learn
  • What an Agent is in Dezifi and how it differs from a Workflow
  • The 9-step builder at a glance
  • When to reach for an Agent vs. a Workflow
  • How Agents tie into Tools, Skills, RAG, Memory, Voice and Guardrails

Agent vs. Workflow

An Agent is a single intelligent worker — one model, one Tool belt, one Guardrail profile. A Workflow orchestrates one or more Agents with branching, loops and human approval. Start with an Agent. Promote to a Workflow when you need explicit control flow.

The 9-step builder

Every Agent is configured the same way. Each step is independent — you can revisit any of them.
  1. 1

    Basic Info

    Name, description, and category (DevOps, Security, Quality, Support, Productivity, Analytics, or Custom).
  2. 2

    LLM Selection

    Pick the model — OpenAI, Anthropic, Google, AWS Bedrock, Azure OpenAI, or a locally hosted model.
  3. 3

    Tool Selection

    Choose which integrations the Agent can call at runtime.
  4. 4

    Skill Selection

    Attach reusable prompts and instruction sets from the Skill library.
  5. 5

    RAG Configuration

    Bind one or more Knowledge Bases for retrieval-augmented responses.
  6. 6

    Memory Configuration

    Pick short-term, long-term, or hybrid memory depending on whether the Agent should remember across Sessions.
  7. 7

    Voice Configuration

    Optionally enable speech-in / speech-out with a configurable STT and TTS provider.
  8. 8

    Guardrails Configuration

    Attach a Guardrail profile to enforce safety, PII and policy checks on every Run.
  9. 9

    Review and Publish

    Confirm the configuration. Publishing creates a versioned Agent that can be tested, exposed as an API, or wired into a Workflow.

When to use an Agent

Reach for a single Agent when the job is conversational, open-ended, or fits a single question/response shape — support triage, code review, document Q&A, internal analytics chat. Reach for a Workflow when the job has fixed steps, multi-Agent handoff, scheduled triggers, or human approvals.

Frequently asked questions

Can one Agent use multiple LLMs?
Each Agent runs on a single primary LLM, but you can route Tools through their own models and chain Agents in a Workflow that mixes providers.
Do I have to pick every step in the builder?
Only Basic Info, LLM Selection, and Review are required. Tools, Skills, RAG, Memory, Voice and Guardrails are all optional — leave them empty and add them later.
How are Agents versioned?
Each Publish action creates a new version. Past versions remain available — you can roll back, compare in Eval, or run A/B traffic splits.
Where do Agents run?
Inside your Workspace. Each tenant is isolated. On-premise and private-cloud deployments execute Agents inside your own infrastructure.