What's new in the 2026 Workflows engine

The 2026 Workflows engine rebuilt Attio’s automation around AI agent steps, native JavaScript, and a credit model where most blocks are free.

The 2026 Workflows engine rebuilt Attio’s automation around AI agent steps, native JavaScript, and a credit model where most blocks are free. Attio shipped it in June 2026. At the same time it closed the previous engine to new creation. You can no longer create new legacy workflows; existing ones keep running, and Attio says sunset notice will come “well in advance.” This is not a facelift. It is a different product. The pricing model is different, and the block library is larger. The builder also treats AI as a real step in the flow, not an add-on.

A GTM (go-to-market) engineer builds the automations that move leads and deals through a CRM. For that work, the practical change is this. Logic you used to push out to Zapier, Make, or Clay can now live inside Attio. It sits right beside the records it reads and writes. An AI agent can research a lead and draft a follow-up. In the same run, an Execute code block can reshape a webhook payload. The read-only, logic-only parts of that run cost nothing.

This page explains what changed since the legacy engine, and how each change affects the way you build. After reading it, you can decide whether to migrate an existing workflow. You can also see which new blocks unlock work you could not do before. And you can think through the new credit model before you ship anything to production.

Running agents as workflow steps

The headline change is that AI agents now run inside a workflow as ordinary steps. An agent can read context, decide, and hand its output to the next block in the same run. The Custom agent block configures a model with an instruction prompt and a set of tools. You can pick GPT, Sonnet, Opus, or Gemini Pro. Its access to your workspace is read-only, gated by the “Run as” permission model. So an agent can read records, notes, emails, and call recordings, but it cannot write or delete on its own. When you want a change saved, you place an explicit write block after it.

A minimal agent step needs three things: a model, a prompt, and the record context it should read.

{
  "block": "custom_agent",
  "model": "Sonnet",
  "run_as": "[e-mail]",
  "instructions": "Summarize the latest three notes on this company and flag any churn risk in one sentence.",
  "context": "{{ trigger.record }}"
}

The agent returns a text result. You reference it downstream as {{ steps.custom_agent.output }} and feed it into an Update record write or a Slack message. The Web agent block is the former “Research record” block rebuilt. You ask a question about a record, and the agent researches across the web to answer it. You can also describe a workflow in plain language in the canvas and let Ask Attio assemble it block by block, live.

Writing custom logic with native code

The Execute code block runs your own JavaScript as a step. This removes the external webhook round-trip that data shaping used to require. You export a main(inputs) function, and whatever it returns becomes the block’s output. Transformations, custom scoring, and payload reshaping now run inline, beside the records they touch. Each run has a 100-second timeout.

export function main(inputs) {
  const domain = inputs.email.split("@")[1];
  const tier = inputs.employees > 500 ? "enterprise" : "mid-market";
  return { domain, tier };
}

Given [e-mail] and an employee count of 1,200, this block returns { "domain": "acme.com", "tier": "enterprise" }. A later block reads that as {{ steps.execute_code.tier }}. The cost of Execute code scales with the work it does. So its credit cost is listed as Variable (complexity-based) rather than a flat per-run figure. For heavier integration logic, the Send HTTP request block pairs naturally with code. It calls an external API, and code parses the response in the same workflow.

Paying for outcomes, not steps

The credit model was rebuilt around one rule. You pay only when a block does something that costs Attio money to run. Triggers, conditions, lookups, and delays are free. Credits are consumed only when a block writes data, sends or receives something externally, or invokes AI. This shift matters most for read-heavy, logic-heavy workflows. They used to charge a credit per step, and now they run their logic for nothing.

Consider a routing workflow. It fires on a Webhook received trigger, reads three records, evaluates two conditions, and ends in a single write. Under the legacy engine, every one of those steps consumed a credit. Under the 2026 engine, only the final write does. So the same workflow costs one credit per run instead of six or seven. AI blocks moved the other way. Instead of a fixed 1 to 10 credits, they now bill by the tokens the model consumes. That makes an AI-heavy workflow variable-cost, and worth metering before you scale it.

One legacy safety valve is gone. Per-workflow credit caps were removed in the new engine. So you can no longer set a hard credit ceiling on a single workflow. Budget at the workspace level and watch run history instead. The full per-block table, plan allowances, and guidance on reasoning about variable cost live on Credits & pricing explained.

Building with the new canvas

The builder gained the quality-of-life features that made the legacy engine tedious for anything beyond a single trigger. A workflow can now start from several events at once. These are called multi-trigger workflows. You resolve whichever event fired using the fallback patterns described in Variables & data passing. You can copy and paste blocks between workflows, and even between workspaces. That turns a tested pattern in one client’s workspace into a starting point in another’s.

In-flight runs are now recoverable. You can retry, pause, and resume them. You can also re-run a failed step without restarting the whole workflow, as covered in Runs, debugging & publishing. Blocks drag in from the sidebar, and HTTP payloads have a visual JSON editor. You can also drop a block straight into a loop. The library itself is extensible. You can ship your own blocks with the App SDK custom blocks workflow. Attio also added ecosystem blocks for Webflow, Notion, Linear, and more.

Deciding whether to migrate

If you have legacy workflows running, they keep running. Attio says sunset notice will come “well in advance.” But new capability and the cheaper pricing exist only in the new engine. Several blocks were renamed, and a few behave differently. So work through Migrate legacy workflows and check the Block changes reference before you rebuild anything.

If you are starting fresh, ignore legacy entirely. Learn the new engine’s Block reference overview. Then learn the handful of core ideas in Choosing the right trigger that everything else hangs off. The judgment call is no longer whether Attio can host a piece of logic. It is whether it should. The Guides section shows where workflows genuinely win against a standalone automation tool.

Sources: Migrate your legacy workflows and the Attio changelog. Last reviewed against the Attio Help Center, June 2026.

See also

• Credits & pricing explained
• Custom agent
• Execute code
• Migrate legacy workflows
• What are Attio Workflows?