Hermes Agent Deep Dive
I went inside Hermes Agent's loop, memory, and context management, then put it to work on a QA plan with 174 business cases.
Singapore Management University · Singapore · 8 June 2026
Overview
I presented this deep dive at the Hermes Agent Singapore Meetup at Singapore Management University.
I wanted to explain how Hermes carries work through a long session: gathering context, calling tools, saving history, and deciding what to keep in the next prompt.
I then compared Hermes and OpenCode as QA runners using gstack. My test plan covered quotations, orders, invoices, payments, calendars, and business rules.
Topics and QA Workflow
Agent mechanics
I traced a message through prompt assembly, tool execution, and session storage, then showed how search and compression make older work reusable.
Automated QA
I ran business scenarios, classified the outcomes, and followed a correction cycle. I kept partial and blocked cases separate from passes.
Architecture and Tools
I covered SQLite session storage and FTS5 search over messages and tool calls. Saved history stays searchable without putting every earlier message back into the active prompt.
For context management, I explained cached prompts, temporary context overlays, and a four phase compression process: prune old tool results, preserve the beginning and recent turns, summarize the middle, and reassemble the transcript.
What I Observed
In my internal evaluation run, Hermes passed 105 of the 174 cases after one autonomous correction cycle. The remaining cases stayed classified as partial, blocked, failed, or not run. Hermes used about 50 percent fewer billable tokens than OpenCode in this comparison.