Artificial Intelligence Daily edition
ARTIFICIAL INTELLIGENCE AGENTS AND MCP
Agent Observability Requires More Than a Trace Dashboard
APM tools log calls, not decisions. Without a record of intent, post-incident reviews for agent activity rely on reconstructions.
An agent executing infrastructure changes produces two distinct trails: the technical trail, which any APM can capture, and the decision trail, which almost no one stores. The former explains what happened; the latter explains why. Teams that only retain the former discover during incident review that they cannot distinguish a tool failure from an error in judgement.
What to Log Besides Traces
- The effective instruction received, along with the system prompt version.
- The tools available at the time, and those that were denied.
- The intermediate plan before the first write operation.
- The identifier for the human or process that authorised the execution.
With these four fields, reconstruction ceases to be a narrative and becomes an audit. Without them, the post-mortem ends with the sentence found in almost every report we have reviewed this year: "the agent acted unexpectedly".
An Operational Baseline
{
"run_id": "01J...",
"actor": "svc-agent-ops",
"authorized_by": "user:elena",
"tools_allowed": ["k8s.read", "k8s.scale"],
"tools_denied": ["k8s.delete"],
"plan_hash": "sha256:9f2c...",
"writes": 3
}Our first agent-driven incident lasted eleven minutes. Understanding it took us nine days.
What to Expect from Providers
Orchestration platforms are beginning to expose the intent log, but with short retention periods and no standard export format. Until a common format exists, it is advisable to duplicate this log in your own data store. The cost is marginal, and its value becomes apparent on the very day it is needed.
Tags
- Agents
- Observability
- Post-mortem
BOLDERROR Daily edition Rubén Campoy