Retry Tax
A system can look cheap per model call and still be expensive per successful task. Retries, fallbacks, human escalation, and tool failures belong in the denominator.
FIELD NOTES / OPERATING OBSERVATIONS
Signals from building enterprise AI.
Concise observations about failure modes, economics, architecture, and control. Each note stands on its own as a practical idea.
Agent Sprawl
When every team can create agents quickly, discovery and reuse become architecture problems. The registry is not the transformation engine, but the engine needs one.
Context Inflation
More context is not free. It raises latency, spend, and sometimes confusion. Context should be treated as a budgeted resource, not an unlimited prompt attachment.
Not Everything Needs an Agent
If a deterministic workflow solves the problem reliably, adding an agent can increase cost and failure surface without increasing business value.
Protocol ≠ Orchestration
MCP can standardize access to capabilities. It does not decide which capability should run, whether it is safe, or whether the result deserves production trust.
Agents Within Agents
Nested agent systems amplify ownership, prompt variance, observability, and evaluation problems. Composition needs lifecycle controls, not only clever routing.
Abstraction Tax
Framework abstractions are useful only when their productivity gains exceed the debugging opacity, dependency surface, and portability cost they introduce.
Cost per Successful Task
Cost per API call is often the wrong unit. The useful denominator is a completed task that meets quality, safety, and latency thresholds.