A practical field guide from Automation Ace.
Short answer: monitor AI agents in Make on four fronts: logs (which tools each run called, with what inputs and outputs), errors (failed and incomplete runs, with alerts), costs (credits and tokens per run), and quality (a weekly human review sample). Log a summary of every run to a table so you can spot trends, and tighten prompts, steps, and tools when numbers drift.
Short answer
- Review run details for tool calls and outputs.
- Alert on errors and repeated failures.
- Track credits and tokens per run and per outcome.
- Review a sample of runs weekly for quality.
What to monitor
| What to monitor | Where to find it | Warning signs |
|---|---|---|
| Runs and outcomes | Scenario execution history and agent run details | Rising failure or escalation rate |
| Tool calls | Run details showing which tool was used, inputs, and outputs | Wrong tool chosen; repeated calls; unexpected writes |
| Errors | Errored runs, incomplete executions, error-handler alerts | Same error repeating; silent failures |
| Token and credit usage | Credit usage views and the token usage summary in the agent output | Cost per run creeping up; long prompts or histories |
| Quality | Your own review sample and feedback | Edits needed on outputs; complaints |
| Latency | Run durations | Slow runs blocking downstream steps |
Logs: see what the agent did
Make's run history shows which tool an agent used, the inputs it sent, and what it got back. That is your first stop when an agent does something unexpected. For trends, also write a compact log row per run:
Log row per agent run (Data store, Airtable, or a database)
run_id, scenario, agent, started_at, duration_s,
input_ref (record id, not full text), outcome (done/escalated/failed),
tools_called (comma list), tokens_in, tokens_out, credits,
error_message, reviewer_flag (ok/needs_fix), notesStore references, not full customer text, when data is sensitive. For logging patterns, see run logs with tables (the same idea applies in Make with data stores).
Errors: catch failures before users do
- Wrap the agent call with an error handler that logs and alerts; see Make error handling.
- Use incomplete executions to retry where appropriate; see incomplete executions.
- Validate the agent's structured output before downstream modules act on it.
- Alert on patterns, not just single failures: a rising escalation rate matters.
Costs: understand and control credit usage
Make's documentation explains that agent runs use credits per operation plus credits based on AI token usage when you use Make's AI provider, and that knowledge queries add credits too; the agent's output includes a token usage summary. With your own AI provider, model costs are billed by that provider. Check the current credit usage docs for exact rates. Ways to reduce cost:
| Cost driver | How to reduce it |
|---|---|
| Large inputs | Send only the fields the agent needs; summarize long text first |
| Long history | Lower max history; avoid passing threads when not needed |
| Many steps | Tighten max steps; give clearer tool descriptions so the agent finishes faster |
| Model choice | Use a smaller model for simple classification jobs |
| Work that does not need AI | Move fixed steps to normal modules and scenarios |
| Knowledge queries | Keep knowledge focused; avoid querying it for every run if not needed |
See why a scenario uses too many operations.
Quality: review a sample every week
- Pick 10 to 20 runs at random plus every escalation.
- Mark each as correct, acceptable, or wrong, and note why.
- Fix the root cause: prompt, tool description, knowledge, or input data.
- Re-run your test set after changes; see versioning and testing prompts.
Build a simple monitoring dashboard
With a run log table, a dashboard can show runs per day, outcome mix, average credits per run, top errors, and reviewer flags. A weekly digest to Slack is often enough; see sending a digest instead of many alerts.
Platform observability: Make Grid and beyond
At Waves '26, Make described a “glass box” approach to agents, with Make Grid as a live map of automations (including agents on other platforms), searchable execution logs, and OpenTelemetry streaming to existing monitoring tools, some of it marked as coming soon. Check what is available on your plan; see what MakeOS is.
A weekly monitoring routine
- Check the dashboard: volume, failures, escalations, cost.
- Review the sample and every escalation.
- Fix the top issue and re-run tests.
- Record changes in a changelog.
Before monitoring, build well: see how to build AI agents in Make and controlling agent permissions.
Frequently asked questions
How do I monitor AI agents in Make?
Review run details to see which tools each run called with what inputs and outputs, alert on errors and repeated failures, track credits and tokens per run, log a summary of each run to a table, and review a sample of runs every week.
How much do Make AI agents cost to run?
Make’s documentation describes agent runs using credits per operation plus credits based on AI token usage when using Make’s AI provider, with knowledge queries adding credits. With your own provider, model costs are billed by that provider. Check Make’s current credit usage docs.
How can I reduce AI agent costs in Make?
Send only needed fields, lower max history and max steps, write clearer tool descriptions, use a smaller model for simple jobs, move fixed steps to normal modules, and keep knowledge queries focused.
Where can I see what a Make AI agent did?
Make’s run history shows which tool the agent used, the inputs it sent, and what it returned. Logging a compact summary of each run to a table helps you see trends over time.
What is Make Grid?
Make Grid is Make’s live map of automations. At Waves ’26 Make described it as including agents on other platforms, alongside searchable logs and OpenTelemetry streaming, some marked as coming soon.
Disclaimer: Make features, module names, plans, and credit usage change over time, and several capabilities announced at Waves '26 (such as MakeOS and Process Intelligence) were in beta or listed as coming soon at the time of writing; confirm current details in the Make Help Center. This article may include links to apps, products, or services; some links may be affiliate links, which means Automation Ace may earn a commission at no extra cost to you.