Claude Code verifies a job-scheduling research report
Borchers reports researching trigger.dev alternatives with Claude Code workers, auditing false claims, and rerunning with verifiers; errors decreased in his own review but did not disappear.
drborchers.com
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Research alternatives to [job scheduler]. Preserve source evidence and intermediate findings. Refetch critical claims before synthesis, flag worker/checker disagreements and separate verified, outdated and unresolved claims.
Reach people exploring AI agents
See placements →What happened
Borchers reports researching trigger.dev alternatives with Claude Code workers, auditing false claims, and rerunning with verifiers; errors decreased in his own review but did not disappear.
Reported by Oliver BorchersHow the agent did it
- Assign specific scheduling-alternative topics to research workers.
- Preserve their findings and source evidence in intermediate files.
- Audit the first report and identify unused verification instructions.
Where a person stepped in
Borchers audits outputs and revises the pipeline. His checks find remaining errors and a verifier-created mistake; the reported error rates are not a controlled benchmark.
Oliver Borchers tests a Claude Code research plugin on alternatives to trigger.dev for job scheduling. Five workers investigate the baseline service, managed platforms, self-hosted queues, Supabase approaches and serverless scheduling, preserving intermediate documents before a separate agent synthesises them.
The first report looks convincing, but a later check of sixty-three claims identifies ten errors and two outdated items. Problems include an invented funding figure and an inverted database-connection claim. Borchers discovers that his verification-method skills were written but never referenced by the agents actually doing the research.
He revises the pipeline so workers produce claim tables, verifier agents refetch sources and the synthesiser sees both findings and verification reports. A rerun has fewer serious fabrications, yet a verifier also introduces a wrong correction to a previously correct download count. The author's before-and-after audits use different claim totals, so they are not a controlled benchmark. The useful artifact is a traced research report with preserved intermediate evidence and confidence notes, plus a concrete demonstration that an agent labelled verifier still needs review. Contradictions between worker and checker should remain inspectable rather than gaining authority simply from their position in the pipeline.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Blog · Oliver Borchers
Frequently asked questions
What failed in the first pipeline?
Verification skills existed in documentation but were not referenced by the research agents.
What changed before the rerun?
Workers added claim tables, verifiers refetched sources and the synthesiser read verification reports before producing the final document.
Did verifiers eliminate false claims?
No. Remaining errors persisted, and one verifier incorrectly changed a worker’s previously correct download figure.
More details
Remaining steps
- Require claim tables and independent source refetching before synthesis.
- Inspect disagreements and the rerun’s remaining errors.
About this account
Claude Code workers investigate multiple alternatives, verifier agents refetch evidence and a synthesiser assembles a report from both findings and checks.
Firsthand account by Oliver Borchers; original publication date verified against Primary blog March 15, 2026. Outcomes are the author’s reports, not independently reproduced. Suggested prompt written by Agents In Use; not the author’s original prompt.
Credit: Oliver Borchers
Source published 15 Mar 2026
Source checked 14 Oct 2026
Tools and topics
Claude Code · Research workers · Verifier agents · Web search · GitHub APIs · Claim tables
Scheduling research · Claim verification · Conflicting corrections
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.