[object Object]

A team running Breeze agents alongside HubSpot workflows treats them as separate tools — agents do their thing in the background, workflows do automation, and reps stitch the outputs together manually. The “Run Agent” workflow action collapses that gap. It is in private beta but already changes how you design automation, because an agent becomes a workflow step you can branch on.

The new primitive

The workflow library gains an action: Run Agent. Configure which agent to invoke, pass context properties, and capture structured outputs that feed downstream steps. The agent’s reply is no longer a side effect — it is data the rest of the workflow uses.

Workflow step: Run Agent
  Agent:        Company Research Agent
  Inputs:       contact.email, contact.company
  Outputs:      industry, employee_count, funding_stage, ai_confidence
  Timeout:      60 seconds
  On error:     skip downstream branch, fallback to manual enrichment

This pattern works across any agent that returns structured output — Customer Agent for triage results, Prospecting Agent for scored leads, Data Agent for cleansed properties.

Pattern: enrich before route

A new contact lands from a form. Without enrichment, lead routing falls back to coarse rules (region only). With agent enrichment in the workflow, routing reads the freshly populated industry and headcount before deciding the assignment:

Workflow: Inbound lead routing
  Trigger: form submission "Demo request"
  Step 1:  Run Agent: Company Research
           outputs: industry, employee_count, funding_stage
  Step 2:  Branch on industry + employee_count
           industry=Healthcare AND employees>500 -> Sarah (Healthcare AE)
           industry=Software AND employees>1000  -> Tom (Enterprise AE)
           else -> SDR queue
  Step 3:  Create task for assigned owner
  Step 4:  Notify owner in Slack with enrichment summary

The previously manual research step becomes zero-touch and routing accuracy lifts measurably.

Pattern: triage before escalate

A support ticket arrives. Customer Agent attempts a first-pass resolution; only unresolved tickets reach a human, and they reach with context:

Workflow: Ticket triage
  Trigger: ticket created on tier-1 pipeline
  Step 1:  Run Agent: Customer Agent
           outputs: resolution_status, intent, summary, suggested_action
  Step 2:  Branch on resolution_status
           = "resolved" -> close ticket, request CSAT
           = "needs_human" -> assign to tier-1 human queue
                              attach summary + suggested_action
           = "needs_engineering" -> escalate to eng queue
                                    attach reproduction steps

Deflection is baked into the workflow rather than living as a separate bot console.

Pattern: enrich before campaign

A new contact gets enrolled in a nurture campaign. Without clean data, personalization tokens break and emails feel generic. Run Data Agent before the first send to fill gaps:

Workflow: Nurture enrollment
  Trigger: contact added to nurture list
  Step 1:  Run Agent: Data Agent
           inputs: existing properties
           outputs: filled_role, filled_industry, filled_company_size
  Step 2:  Wait 1 hour (let agent results settle)
  Step 3:  Send email 1 of nurture sequence
  Step 4:  Branch on engagement

Error handling

Agents can time out, fail, or return low-confidence results. Build branches for each:

// Custom code action wrapping agent output
const agentOutput = event.inputFields.agent_output;
if (!agentOutput || agentOutput.confidence < 0.6) {
  return { outputFields: { skip_downstream: true, reason: "low_confidence" } };
}
return { outputFields: { skip_downstream: false, ...agentOutput } };

A “low confidence” branch can route to a human review queue rather than acting on bad data.

Cost control

Outcome-based pricing on agents means each workflow run that calls an agent has a unit cost. Set spend alerts and audit weekly:

- Daily agent invocation count per workflow
- Spend per workflow per week
- Average confidence score per agent run
- Skipped-due-to-low-confidence rate

A workflow burning agent calls on records that always score low confidence is a candidate to add a pre-filter before the agent step.

Private beta caveats

Private beta means contracts can change. Avoid building load-bearing automation on Run Agent until GA. Use it now for non-critical enrichment, triage, and personalization where a temporary outage is recoverable.

What to do this week

Identify one manual research or triage step in your highest-volume workflow, prototype the Run Agent equivalent in test mode, and instrument cost plus confidence metrics before requesting beta access if you do not have it already.

[object Object]
Share