Notifications
Clear all
Topic starter
18/08/2026 2:39 pm
TL;DR: AI agents investigate production issues better when they can iterate freely over observability data, because fixed APIs and limited context windows constrain hypothesis testing more than model quality or prompt design, according to Edge Delta. The practical shift is toward sandboxed execution, unified schemas, and tighter control over tool-driven investigation workflows.
NHIMG editorial — based on content published by Edge Delta: Agentic observability and investigation freedom in AI Teammates
Questions worth separating out
Q: How should security teams govern agentic AI as it moves into production?
A: Security teams should govern agentic AI as a class of non-human identity, not as a generic application feature.
Q: What breaks when observability agents are limited to fixed tools and schemas?
A: They can only ask the questions the tools were built to answer, which pushes them into chained calls, partial results, and brittle normalisation.
Q: How do you know if sandboxed agent investigations are actually working?
A: Look for fewer malformed queries, lower tool-call counts per case, and higher ratios of raw data processed outside the model context.
Practitioner guidance
- Define agent investigation identities Assign each AI investigation workflow a distinct non-human identity with scoped permissions for telemetry access, code execution, and remediation actions.
- Constrain sandboxed execution paths Permit Python, Bash, and repository inspection only inside isolated sandboxes with no implicit access to adjacent systems or long-lived credentials.
- Standardise telemetry schemas Reduce schema drift across logs, metrics, and traces by aligning field names and query language semantics before enabling agent-led investigations.
What's in the full article
Edge Delta's full article covers the operational detail this post intentionally leaves for the source:
- How the AI Teammates workflow uses sandboxed Python and Bash to process raw telemetry outside the model context
- Why tool-call counts, malformed query rates, and sandbox offload ratios are the practical metrics the vendor tracks
- How a unified telemetry schema changes query behaviour across logs, metrics, traces, and code inspection
- What the platform architecture means for teams comparing agent-led investigation models in practice
👉 Read Edge Delta's analysis of agentic observability and sandboxed investigation workflows →
Agentic observability and sandboxed iteration: are your controls keeping up?
Explore further
18/08/2026 3:24 pm
Agentic observability creates a new non-human identity governance problem: the agent is not just reading data, it is acting inside a scoped runtime with query and execution privileges. That means the relevant governance question is no longer only what the agent can see, but what it can do with what it sees. For IAM and PAM teams, this looks like ephemeral, task-scoped access control applied to investigative workloads, with the same lifecycle discipline expected for other NHIs.
A question worth separating out:
Q: What accountability controls should exist before AI agents can run analysis on telemetry?
A: Owners should define who approves access, who reviews outputs, and who can revoke the agent’s permissions when the investigation ends. If the agent can reach codebases or remediation paths, those approvals should map to privileged-access governance, with evidence of scope, timing, and offboarding retained for audit and incident review.
👉 Read our full editorial: Agentic observability depends on iteration, not just model quality
