An agent that watches 50+ automation scenarios, and what it taught me
The agent has applied one fix on its own so far. The more useful result is how rarely it needs to wake Claude at all.
A consultancy runs its newsletters, webinars, bookings and academy on a Make account with more than 50 scenarios. When one of them fails at night, nobody notices until a customer writes. I built an agent that notices first, diagnoses the cause and repairs what can safely be undone.
The problem that started it
Make parks nothing by default. One malformed record deactivated its scenario, and every webhook call behind it waited in the queue. In one case nine records sat there for five days before anyone saw it. The cost was small each time and the pattern was the same each time: a bad input, a stopped scenario, a silent backlog.
How the agent is built
A short script runs every 15 minutes and does no analysis. It counts failed runs and parked bundles through the Make API. When the count is zero, it writes one line into a run log and stops. No Claude call happens, so a quiet quarter hour costs nothing in tokens.
Only when the script finds something does it start Claude with a written routine. Claude reads the error, finds the module that failed, classifies the fix and routes it. A change that can be undone, such as correcting a field mapping or an endpoint URL, is applied by the agent, which then reactivates the scenario and checks that Make still accepts it. A change that touches data waits for my approval. Anything the routine does not recognise is refused by default.
- Every incident is posted to the team chat, with the diagnosis.
- A single flag switches the agent to read-only. The first weeks ran in that shadow mode.
- The agent keeps its own state in two Make data stores, and it never runs twice in parallel.
What the run log shows
Between 31 August and 24 October the check ran 4,792 times on my Mac. In 4,710 of those runs it found nothing and wrote its line without a Claude call. The other 82 broke off: 64 during the scan, 18 during the Claude call. The whole log, from 20 July to 24 October, records 101 incidents and one fix that the agent applied on its own, on 1 August. That is a modest number, and I report it as it is. Most failures were transient and Make retries those by itself, so the agent's real work is the smaller set of errors that Make gives up on.
Most nights the best thing the agent does is nothing, and it does it for free.
Three things that went wrong
First, the agent did not run at all for three weeks. I had built it as a scheduled Claude task, and the Make connector I used while building does not exist in headless runs. Every tick failed. The fix was a small helper that talks to the Make REST API directly, with the token in the system keychain. After that the connector was no longer needed.
Second, updating a scenario through the API deactivates it. A fix that lands on a disabled scenario is worse than the bug. The executor now reactivates the scenario and verifies it, and a dry run on a throwaway scenario caught this before it touched anything real.
Third, my own bulk script died at the first rate-limit error. The Make API allows about 60 requests a minute for the whole organisation, and a full scan of 58 webhook scenarios burns that in seconds. Bulk jobs now pace themselves and back off.
Parked bundles and a blind spot
To stop one bad record from freezing a queue, every webhook scenario now runs with a dead-letter queue. The bad bundle is parked and the rest keeps flowing. The side effect is that a parked failure counts as a warning, not an error, so a log scan alone no longer sees it. The scan therefore reads the parked-bundle count as a second source. Resolving parked records stays manual on purpose, because a retry replays the same payload.
Since 25 October the check runs as a systemd timer on my own server in Vienna. The Mac job is stopped, because two instances would write to the same stores.
What carries over
Give an agent a check that costs nothing, wake the expensive part only on evidence, and write down which fixes it may apply alone. The rest of the design follows from those three rules. The details of the platform it watches are in the case study on the automation platform architecture.

