Just tried to automate a security review by chaining a code scanner with a report generator. The concept is solid, but the failure modes are brutal.
First agent fails or times out? The whole chain collapses silently. No retries, no partial output, no dead-letter queue. You get a generic "Something went wrong" and have to trace through each agent's logs manually.
Key issues:
* No state persistence between agents. If agent two fails, you lose agent one's output.
* Error propagation is non-existent. You can't easily implement a fallback agent.
* Debugging requires you to simulate the exact chain state, which is tedious.
Example of a fragile chain definition:
```yaml
agents:
- name: vulnerability-scanner
task: "scan $TARGET for CVE-2024-*"
- name: report-builder
task: "format ${vulnerability-scanner.output} into markdown"
# If scanner outputs an unexpected format, report-builder just crashes.
```
Until they add proper error hooks and state management, running anything mission-critical in a chain is a risk. Stick to single, monitored agents for now.
-dk
Trust but verify, then don't trust.
Yep, seen this exact pattern with CI pipelines. If a job fails, its artifacts are often gone unless you explicitly persist them.
The workaround I use: wrap each agent in a script step that catches the exit code and dumps its output to a known file before failing. Then the next agent can at least try to pick up the partial file.
But it's a hack, not a solution. You end up rebuilding a queue.
Ship fast, review slower