Hermes Agent Deep Dive
I showed software teams how Hermes Agent keeps track of a long testing task, then examined whether it could work through failures and return to unfinished checks.
Singapore Management University (SMU) · Singapore · 8 June 2026
Keeping a Long Test Plan on Track
A coding assistant doing extended testing needs more than the latest instruction. It needs to remember the plan, what has already failed, what was changed, and what to try next. At Singapore Management University, I explained how Hermes retains that working history without letting old output overwhelm the current task.
I put those ideas to work in an internal comparison with OpenCode, testing software that handles quotations, orders, invoices, payments, and calendars. The talk follows both correction and unfinished work: Hermes recovered from a stuck loop, but some checks remained partial, blocked, failed, or unrun. The comparison concerns that internal workload, not a general ranking of coding assistants.
Managing a Long Testing Run
Tool Results Become Working History
The agent loop gathered the user message and prompt context, executed tool calls, and appended their results to history before continuing. SQLite retained session metadata, messages, model configuration, and session lineage. FTS5 made past content and tool calls searchable, so earlier work could be recalled rather than relying only on the latest exchange.
Assembling and Compressing Context
Cached prompt inputs were separate from temporary instructions, prefill messages, session overlays, and recalled memory added for the current API call. Durable history and the context sent to a model are therefore different layers: retaining the record does not require sending every old tool result on every turn.
In the architecture presented, session hygiene used an 85 percent context threshold before the agent started; compression inside the loop used 50 percent. Compression pruned old tool results, preserved initial context and recent turns, summarized the middle, and reassembled the transcript. This kept the starting task and recent work available while reducing the accumulated output.
Correction Across 174 Test Cases
I ran Hermes and OpenCode concurrently with gstack against 174 test cases covering quotations, orders, invoices, payments, calendars, and business rules. The first autonomous run produced 56 passes for Hermes and 36 for OpenCode.
After escaping a stuck loop and completing an autonomous correction cycle, Hermes recorded 105 passes, 44 partial results, 21 blocked checks, two failures, and two not run. In this internal workload comparison, Hermes consumed 746,252 billable tokens against OpenCode's 1,500,687. The counts and token use describe this measured run, not expected results for other projects.