Golden Paths for AI Agents: Redesigning Platform Workflows for Non-Human Users

Platform engineering built its golden paths for humans: a developer opens a ticket, picks a workflow, follows steps, and ships. The golden path concept assumes someone is making decisions at each stage. Agents don’t. They execute probabilistically, autonomously, and at machine speed — and the platform infrastructure built around human judgment breaks down when the user has none.

The Framework

Datadog’s analysis identifies the question platform teams have stopped asking correctly: not “should agents use our platform?” but “how must our platform change for agents to use it safely?” The answer requires rethinking three core assumptions that underlie every conventional golden path.

Execution Patterns Must Match the Workload

Human developers work synchronously within tight time windows. Agents don’t share that constraint, which means a single golden path cannot serve all agent workloads.

Real-time conversational agents require low-latency synchronous execution — seconds, not minutes. Long-running proactive agents need durable, resumable execution that survives infrastructure restarts and maintains state across sessions. Code-executing agents require isolated sandboxes with controlled filesystem, network, and resource access.

A platform that forces all three into the same workflow has designed for the wrong user. The practical requirement: define execution profiles for each agent type and route them to purpose-built infrastructure rather than adapting general-purpose paths.

Deterministic Controls Must Sit Between Probabilistic Outputs

Agents diagnose problems and generate solutions probabilistically. Platform teams are responsible for deciding whether those outputs reach production — and that decision cannot be delegated to the agent’s own judgment.

Datadog’s framework uses what it calls “hybrid golden paths”: agents perform probabilistic work (code generation, diagnosis, recommendation), and deterministic gates — automated tests, security scans, approval requirements — control whether that work proceeds. The agent cannot bypass these gates; the platform enforces them structurally.

Datadog’s AI Guard implementation demonstrates this concretely: tool calls are evaluated against security policies before execution, with enforcement occurring outside the agent’s own decision-making. The audit trail records both the agent as actor and the delegating user as principal, creating accountability without requiring the agent to self-report.

Authorization Must Happen Before Execution

Human platform users are authorized at login and act under that authorization. Agents require a different model: each run must be authorized before it starts, not assumed from prior access.

Datadog calls this “dispatch” — converting a signal into a bounded, authorized agent run before execution begins. Credentials issued at dispatch are scoped to specific components and expire when the task completes. The same agent software can operate with different authority levels across use cases without requiring code changes.

Retry budgets replace unlimited correction loops: an agent fixing code might get three attempts or a defined token budget, not an open-ended loop. The platform defines terminal states (succeeded, failed, cancelled, expired, approval denied, escalated) and specifies rollback behavior — because the platform, not the agent, is accountable for the system’s state.

Where It Doesn’t Apply

Golden paths work when the workflow is well-understood and the acceptable variations are bounded. When agents are operating in genuinely novel territory — new codebases, novel failure modes, unexplored integrations — the golden path concept reaches its limits. The hybrid model Datadog describes requires that deterministic controls themselves be correctly specified. An automated test suite that doesn’t cover the relevant behavior is not a safety gate; it’s performance of governance without its substance.

Source: Datadog

Content created with AI assistance and reviewed for accuracy.