All posts
Outbound infra·Jun 9, 2026·4 min read

We benchmarked agent fanout against our SDK pipeline. The SDK ran 1,165% faster.

The Verbiflow teamBy The Verbiflow team

There are two reasonable ways to run batch outbound work from a coding agent. You can have Claude Code write a pipeline using the Verbiflow SDK and run it as one job. Or you can have Claude Code dispatch a fleet of subagents and let each subagent do one slice. We ran the same task both ways: find every security certification claimed on the trust pages of 10 well-known SaaS companies, with the same SDK, model, and task. Here’s what we saw.

Wall clock

Total time to extract certifications across 10 companies
Claude Code + Verbiflow SDK pipeline
31s
Claude Code + 10 subagents
392s
The SDK pipeline ran 12.6× faster (1,165% faster in speed terms, 92% less wall clock). Same AI model, same SDK calls, same 10 companies.

Retries needed

Retries needed before all 10 companies completed
Claude Code + Verbiflow SDK pipeline
0
Claude Code + 10 subagents
4
3 of the 10 subagents refused the first dispatch as “suspected prompt injection.” Same prompt, same workload as the 7 that ran it without issue.
The SDK pipeline kept the repeated work in one batch. Subagent fanout split it into isolated agent loops, which added retries.

Same answer

Security certifications found across all 10 companies
Claude Code + Verbiflow SDK pipeline
85
Claude Code + 10 subagents
87
Within run-to-run search noise. The 2-cert delta is a slightly better trust-page URL drawn by the subagent run on Figma and Snowflake. Re-run the SDK pipeline and the gap can flip directions.

What this benchmark actually shows

This is not an argument against agents or subagents. If you are analyzing one company, asking an agent to do it directly is probably fine. The setup cost of building a pipeline may not be worth it for a one-off task.

But outbound work is rarely one company. Once the same operation needs to run across 10, 100, or 1,000 accounts, the tradeoff flips. The pipeline cost gets paid once, then every additional company gets cheaper, faster, and more reliable.

The subagent version treated 10 companies like 10 separate jobs. The SDK pipeline treated them like one batch job.

The principle

The important part is that the SDK pipeline is not “non-AI.” It still uses LLM calls for extraction, classification, and reasoning where needed. The difference is that those LLM calls run inside a structured pipeline with concurrency, caching, retries, and shared state instead of inside 10 separate agent loops.

Use the agent to design and adjust the workflow. Use the SDK pipeline to run that workflow at scale, including the LLM steps.