Ask someone why their agent system is slow and they will blame the model. The telemetry says otherwise: in my own measured runs, most of the cost and latency accumulated around the model, not inside it. This article names the mechanisms.
The mechanisms
Every hop resends context. When a planner hands work to a worker, the worker receives its instructions again, plus enough history to act. The reviewer reads the output plus the original context. Each hop re-pays for tokens the previous hop already spent.
Oversight costs more than work. Planning tokens, review tokens, and polling waits are all real spend. In the one-agent-two-models comparison — the same tasks through both paths, with per-call traces — oversight across four orchestration layers exceeded the tokens spent doing the actual task. The measured multiplier behind this is in the cost-machine sequence below.
Failures are cumulative, not isolated. A retry is cheap; a retry culture is not. Inbox errors, browser reconnects, tool-output bloat, and stale duplicate workers each add a little. Together they turn a two-minute task into a twenty-minute session.
Interruptions and inference-vs-tool bottlenecks. A cancelled task still bills its tokens. And long tool calls (a browser action, a file write) block the inference loop, so wall-clock time diverges from token cost — which is why "it took ages" and "it cost a lot" are different measurements.
A request through the cost machine
The attribution trap (canonical)
This section is the canonical description of the cumulative-versus-task and unknown-versus-zero problems. Other articles link here rather than re-deriving them.
Cumulative-versus-task metrics and unknown-versus-zero are the two quiet ways dashboards lie. Fixing them is what made honest optimisation possible at all.
What to do about it
The structural fixes are boring and effective: fewer hops (delete layers that don't repay their coordination cost), task-scoped metrics (attribute tokens to the task, not the lifetime counter), bounded tool output (truncate what you would never read), and a stale-worker reaper. The routing rule — cheap workhorse by default, frontier for genuinely hard work — is documented in the routing benchmark article.
Why this matters outside the terminal
"It took ages" and "it cost a lot" are different measurements — and the same trap exists in marketing data, where platform latency and platform spend get conflated. The habit of separating them transfers directly.
Where to go next
- The multi-agent journey — the architectural decisions this evidence produced
- The agent telemetry stack — how to measure your own version of this