Agentic System Design

"The last mile often becomes most of the journey — moving from prototype to production requires robust error handling, full production tracing for non-deterministic debugging, and careful state management." — Anthropic Engineering

When a worker fails in a multi-agent system, it rarely looks like a failure from the outside. It returns a structurally correct response — right format, plausible values — and the orchestrator synthesizes it with the other workers. No error. No log entry. A confident wrong answer reaches your user.

"We embedded scaling rules in prompts to guide effort budgeting, and developed detailed task descriptions with clear boundaries and output formats to prevent duplication." — Anthropic Engineering

The pattern works when task boundaries are explicit and non-overlapping — each worker owns one domain, the orchestrator merges distinct outputs. Boundary blur is the failure mode: two workers in the same domain return conflicting outputs, and the orchestrator must reconcile a contradiction it was never designed to handle.

"Figuring out what parts of our system could be pared down was not obvious from looking at the logs." — Fractional AI Engineering

Every tool call adds tokens. Quality degrades non-linearly as context fills — not as an error, as slightly worse answers you can't see. Fractional AI only discovered their main cost driver after building a profiler: get_all_fields was consuming most of the context because it returned a full data dump when the agent needed specific fields.

"Parallel tool calling — running 3+ tools simultaneously — reduced complex research query time by up to 90%." — Anthropic Engineering

Sequential chains tasks: one failure aborts the run. Parallel runs independent tasks simultaneously — failures are contained, and latency collapses to the slowest single task. Anthropic's production research system reported 90% reduction in complex query time from this switch alone.

"Prefer reversible over irreversible actions. Request only necessary permissions. Err on the side of doing less and confirming with users when uncertain about intended scope." — Anthropic Engineering

Without explicit constraints, an agent executes a DELETE the same way it executes a READ — immediately, no pause. Three rules: prefer reversible actions, request only the access needed for this step, and pause for human confirmation before anything that can't be undone. Fractional AI's Peter the Safety Agent never finalizes an assessment without expert approval — that one design decision made it deployable in a regulated rail industry context.

"We learned that prompt engineering becomes significantly more complex in multi-agent environments. Early iterations suffered from agents spawning excessive subagents and failing to coordinate effectively." — Anthropic Engineering

A prototype validates the concept on inputs you chose. Production is different: messy inputs, intermittent tool failures, context degrading silently, costs you never profiled, deploys disrupting running agents. Each stage introduces a failure class that wasn't visible at the previous one — and none of it is a model problem.