AI Agents Are Replacing Traditional Developer Platforms—But They Need Strong Guardrails

Developers increasingly rely on AI-powered agents to automate platform tasks, marking a shift away from traditional tools like Backstage. These agents combine user input with system context and route it to a language model, which can either respond directly or invoke tools to accomplish work. The approach depends on semantic search to retrieve relevant information from scattered sources—codebases, pull request discussions, team chat logs, tickets, and design documentation.

How Agents Use Context to Take Action#

Semantic search allows an agent to locate the information it needs in the moment, drawing from diverse knowledge sources. As one expert explained at KubeCon & CloudNativeCon Europe, teams embed everything from Git repositories and Slack conversations to Jira tickets and meeting transcripts, capturing the unwritten conventions of how they prefer to operate.

The challenge is that language models can execute anything a developer requests—or anything they infer from a vague prompt. Without proper guardrails, agents risk making unintended changes. Smart teams are building in controls that either block certain actions outright or require explicit approval before proceeding.

Observability as the Key to Safety#

Traditional debugging doesn't work with agents: running the same question twice produces different execution paths, making trace data the only reliable record of what happened. Logging, metrics, and traces offer visibility into:

• Which tools the agent selected and in what order
• Token consumption and cost associated with each decision
• Retrieval quality, showing which chunks of indexed information the agent actually used
• Audit trails for compliance and accountability when agents modify production systems

OpenTelemetry has introduced semantic conventions specifically for generative AI workloads, standardizing how to capture model calls, token counts, and tool invocations. This data flows to observability platforms—both open-source options like Jaeger, Tempo, Prometheus, and Loki, as well as commercial vendors including Datadog, Honeycomb, Dynatrace, and Elastic.

Getting Semantic Search Right#

The quality of an agent's answers hinges less on the embedding model or vector database than on the data fed into it and how it's structured. Common pitfalls include relying too heavily on outdated documentation, embedding entire documents as single vectors (creating vague, imprecise results), and failing to refresh the index as information changes. The fix: split large documents into self-contained chunks, preserve metadata about source and ownership, run continuous ingestion pipelines, and avoid indexing live state that changes by the second. Instead, give agents a tool to fetch current cluster state on demand.