A team treats Breeze agents and HubSpot workflows as separate worlds — agents do their thing in a dashboard, workflows do automation, and a rep stitches the two together by hand. The Run Agent workflow action collapses that gap. It is in private beta, and the configuration patterns that work are the ones that treat the agent like any other workflow step that can fail.
What the action does
The workflow library gains a new action called Run Agent. You pick which agent to invoke, pass context properties, configure timeout, and capture structured outputs that downstream steps reference like any other property.
Action: Run Agent
Agent: Company Research Agent
Inputs: contact.email, contact.companyname
Outputs: industry, employee_count, funding_stage, ai_confidence
Timeout: 60 seconds
On error: skip downstream branch (set agent_skipped = true)
The output is structured. Downstream steps branch on output values rather than parsing free text.
Configuration patterns
Property mapping:
Pass only what the agent needs (smaller context = faster + cheaper)
Bind outputs to typed properties (text, number, dropdown)
Default values for nullable outputs
Confidence score as a separate output for branching
Error handling:
Timeout: skip and flag for human review
Low confidence (<0.6): route to manual queue
API failure: retry once, then fallback workflow
Rate limit: defer 60 seconds, retry once
A workflow without explicit error handling on agent failure stalls silently. Always define the fallback.
Pattern: enrich before route
A common high-value use:
Workflow: Inbound demo request routing
Trigger: Form submission "Request demo"
Step 1: Run Agent: Company Research
Inputs: contact.email, contact.companyname
Outputs: industry, employees, funding, confidence
Step 2: If confidence < 0.5: route to SDR for manual research
Step 3: Branch on industry + employees:
Healthcare AND >500 -> Sarah (Healthcare AE)
Software AND >1000 -> Tom (Enterprise AE)
Else -> SDR queue
Step 4: Notify owner with enrichment summary in Slack
Routing accuracy improves measurably and the manual research step disappears.
Pattern: triage before escalate
Workflow: Support ticket triage
Trigger: Ticket created on tier-1 pipeline
Step 1: Run Agent: Customer Agent
Inputs: ticket.subject, ticket.body, contact.history
Outputs: status, intent, summary, suggested_action
Step 2: Branch on status:
"resolved" -> close ticket, send CSAT
"needs_human" -> assign to tier-1 queue with summary
"needs_eng" -> escalate to engineering queue
The human picks up tickets the agent could not handle, with the agent’s summary attached so they do not start from scratch.
Output binding gotchas
- Output names are case-sensitive
- Renaming an output silently breaks downstream references
- Untyped outputs fall back to text (not great for branching)
- Null handling: define what happens when output is absent
- Long string outputs may truncate in property storage
Document output schemas as part of the workflow design. Treat them like API contracts.
Cost monitoring
Outcome-based pricing on agents means each invocation has a unit cost. A workflow that calls an agent on every record can run up significant spend:
Daily metrics per workflow:
- Agent invocations
- Average tokens per invocation
- Cost per invocation
- Skipped due to low confidence
- Failed due to error
Pre-filter agent calls when possible. A workflow that runs Company Research only on contacts missing industry data costs a fraction of one that runs it on every contact.
Private beta caveats
Private beta means contracts can change. Avoid building load-bearing automation on Run Agent until it goes GA. Use it for non-critical enrichment, triage, and personalization where a temporary outage is recoverable. Document fallback paths so a beta-feature regression does not cascade into customer-facing failure.
Migration plan when GA arrives
- Audit workflows still using Run Agent
- Verify configuration matches GA action signature
- Re-test in sandbox with GA version
- Update fallback paths if error contracts changed
- Promote to production with monitoring
What to do this week
Pick one manual research or triage step in a high-volume workflow, prototype the Run Agent equivalent in test mode, define explicit error handling for timeout and low confidence, and instrument cost plus confidence metrics before requesting wider beta access.