Compare three overnight agent workflow architectures
At a glance
A builder compares a single agent, a multi-agent handoff, and a deterministic script over recurring real-work tasks. This workflow is listed under agent architecture, workflow automation, evaluation, context management.
- Agents / tools
- Claude Sonnet, Claude Opus, deterministic workflow script, append-only handoff file
Building a tool for these workflows? See weekly rates.
What they did
1) Use representative work such as code reviews, multi-file data processing, reports from raw notes, and email drafts from meeting transcripts. 2) Run three architectures for two weeks each: a single Sonnet agent with four skills and a sequential queue; an Opus orchestrator routing work to three Sonnet workers through an append-only handoff file; and a deterministic script that owns file operations, sequencing, and state while calling an agent only for judgment tasks such as classification, summarization, or drafting. 3) Log each output, compare it with what the task required, and timestamp failures. 4) Use the comparison to decide which parts should remain deterministic and which should be handed to an agent.
How it went
The author reports that only the script-led setup consistently produced usable work. They say the single-agent run began mixing task context after about four to five hours, while multi-agent handoffs lost an estimated 20 to 30 percent of nuance.
Worth knowing
This is the author's self-reported comparison; the post does not provide controlled task counts or a benchmark dataset.