Automated software testing
Work through a test plan for quotations, orders, invoices, payments and calendars while retaining results, corrections and unfinished checks.
Singapore Management University · Singapore · 8 June 2026
Keeping a long test plan on track
Hermes Agent ran tools and appended their results to working history, making corrections while retaining what it had already tested.
Sessions and searchable results
SQLite retained sessions, messages, model configuration and session lineage. FTS5 supported recall of earlier content and tool calls.
Compressing model context
Prompt caching and context compression reduced repeated input while preserving the starting task and recent results. Compression pruned old tool output and summarised the middle of the transcript sent to the model. Session history remained in SQLite.
Corrections and unfinished checks
The internal workload contained 174 test cases. After recovery from a stuck loop and an autonomous correction cycle, Hermes recorded 105 passes, 44 partial results, 21 blocked checks, two failures and two not run.