Video: "NEW Prime Agent BEATS Hermes Agent!?" by Julian Goldie on YouTube.

The core difference most comparisons miss

Prime Agent, released 6 August 2026 by Prime Intellect, keeps a self-editing notebook of its own problem-solving steps. Every 25 turns or so, it reviews what it did, makes small evidence-backed edits to its notes, and starts the next session with that accumulated context. The result is an agent that gets progressively better at the specific kind of task you keep giving it — debugging the same codebase, for instance, or handling a recurring data pipeline problem.

Hermes works differently. It does not keep a solo notebook. Instead, it reads from and writes to a shared Obsidian vault that all your agents — Hermes, Claude, and any others in the stack — can access. The vault holds business rules, project history, style guides, and decisions made in previous sessions. Teaching the vault once means every agent in the system benefits, not just the one that learned it.

That is the real distinction. Prime Agent is a deep solo worker that improves at one job. Hermes is the connected operator that works best as part of a wider system where shared memory matters more than individual depth.

What the benchmark actually tested

Julian Goldie ran both agents on coding tasks: writing functions, debugging errors, and building small automation scripts. On isolated tasks with no prior context, Prime Agent's self-editing notebook gave it an advantage after the first few sessions — it had better notes from earlier runs to draw on, whereas Hermes was starting fresh each time in the absence of a pre-populated vault.

That is a reasonable comparison, but it is also a setup that favours Prime Agent by design. If the same test had run with a well-configured Hermes vault containing relevant business context — the kind of context a real business accumulates over months of use — the results would likely look different. In practice, Hermes is rarely used for isolated single-agent coding tasks the way the benchmark frames it.

Where Prime Agent genuinely wins

For a solo developer or a small team working on a well-defined recurring project — maintaining a bespoke internal tool, running regular data transformations, iterating on a specific codebase — Prime Agent's self-editing notebook is a real advantage. The agent builds up a working record of how problems in your specific environment tend to go wrong and what fixes work. That record is private to that agent and that project.

It also runs locally and is fully open source under an MIT licence, so there are no usage costs once you have the compute and no question of what the tool is sending to a remote server. For businesses with sensitive code, that matters.

Current limitation: Prime Agent supports Python only at launch. No TypeScript, no Go, no mixed-language projects for now.

Where Hermes remains the better choice

If your automation work spans multiple tools — content pipelines, SEO workflows, customer communications, internal reporting — Hermes's ability to share memory across the whole stack is worth more than any single agent's individual improvement rate. A Hermes vault that contains your brand voice, your product lines, your keyword targets, and your decision history is something Prime Agent cannot replicate by design.

Julian Goldie's own recommendation, stated in the video, is to run both for different use cases: Prime Agent for deep, contained coding work; Hermes for the connected business automation stack. That is sensible advice, not fence-sitting.

Where this connects to NordSys

Choosing the right agent — or the right combination of agents — for your automation work depends on what you are actually automating and whether shared memory between tasks matters for your business. Getting that decision wrong means months of work on a tool that does not fit the problem. Our Programming service covers AI coding agent setup, from choosing the right tool to configuring it for your specific stack and project context.

See our Programming service →