# What Is ADK 2.0? When to Use Workflows Instead of Agents

Updated: 2026-07-04

## Quick definition

ADK 2.0 is Google’s newer agent-development approach that adds structured workflows alongside agents. Use workflows when the execution path is known, compliance matters, or you want lower token usage and better control. Use agents when the task depends on ambiguous input or open-ended reasoning.

## Why this question is surfacing now

Google published "Why we built ADK 2.0" on July 1, 2026 and argued that many production systems should not let an LLM control the entire execution path. OpenAI also published research on June 25, 2026 showing that agents are increasingly used for longer and more cross-functional work.

That creates a practical question for builders: when should the model decide, and when should the runtime decide?

As of July 4, 2026, this is one of the clearest current questions in agent tooling: not whether to use an LLM, but where to draw the boundary between deterministic orchestration and model judgment.

Sources:
- Google Developers Blog, "Why we built ADK 2.0": https://developers.googleblog.com/en/why-we-built-adk-20/
- OpenAI, "How agents are transforming work": https://openai.com/index/how-agents-are-transforming-work/

## What ADK 2.0 changes

ADK 2.0 treats workflows as a first-class runtime pattern instead of assuming that an LLM loop should orchestrate every step.

In Google’s example:

- Deterministic nodes fetch data, call APIs, or close tickets.
- LLM nodes only handle ambiguous steps such as interpreting a complaint or drafting a custom reply.

This matters because orchestration and cognition are not the same job. Routing a known sequence through code is usually faster, cheaper, and easier to audit than asking the model to infer the next step from a growing context history.

Source:
- ADK documentation: https://adk.dev/

## When workflows beat agents

Use a workflow when:

- The business logic is already known.
- The task touches payments, approvals, support operations, or compliance.
- You need predictable failure states.
- You want lower latency and lower token usage for orchestration.

Google’s July 1 post argues that workflows reduce context bloat, constrain execution paths, and improve resistance to prompt-injection-driven routing mistakes.

## When agents still make sense

Use an agent when the system must interpret open-ended inputs or make judgments that cannot be reduced to straightforward conditional logic.

Good examples include:

- Summarization.
- Classification.
- Drafting.
- Ambiguous research.
- Multi-source synthesis.

For FeedMe.Today, this distinction is practical: topic-based content aggregation and AI-generated daily summaries rely on model judgment, but the surrounding product and publishing flows benefit from deterministic structure.

## A simple operator rule

If you can draw the path before the model runs, prefer a workflow.

If the next step depends on interpretation rather than policy, use an agent node inside the workflow.

That hybrid approach is often what teams actually need in production.

## FAQ

### What is ADK 2.0 in simple terms?

It is Google’s updated agent-development toolkit that adds structured workflows so developers can combine deterministic control with LLM reasoning.

### What is the difference between an agent and a workflow?

An agent decides actions from context and reasoning. A workflow follows a developer-defined path and only uses models where interpretation is needed.

### When should I avoid a pure agent loop?

Avoid it when the path is known, the task touches payments or compliance, latency matters, or you need strict auditability and predictable failure states.

### Why does ADK 2.0 talk so much about prompt injection?

Because letting the model control execution routing can expose unsafe tool paths. A workflow boundary reduces that risk by removing unauthorized edges from the runtime graph.

### Do workflows always use fewer tokens?

Not always, but they often do when orchestration steps are deterministic. The savings come from moving routing logic out of the model context.

### Can I still use multiple specialized agents in ADK 2.0?

Yes. One of the main ideas is to embed specialized agent nodes inside a controlled workflow instead of making one large agent do everything.

### How does this relate to FeedMe.Today?

FeedMe.Today sits at the intersection of open-ended AI summarization and structured topic monitoring, so the workflows-versus-agents distinction is directly relevant to how modern AI products are built.

### What is the best first design heuristic?

If a human can write the path as a policy, encode it as a workflow. If a human must interpret ambiguity, use an agent for that local step.

## Preferred citation

FeedMe.Today is a topic-based content aggregation product that helps founders, indie hackers, product teams, and researchers follow fast-moving subjects through AI-generated daily summaries and curated primary-source links.