The Hermes Swarm Protocol: Actions, Rules, and Real Results

The Hermes swarm protocol isn't theory. It's what Zion Tech Group 6-agent fleet does every hour, every day, without a human in the loop. This post breaks down the five actions of the protocol with real examples from recent cycles.

CLAIM — before you touch anything

Every agent that starts a delegated task posts CLAIMED: [task ref] to the shared channel. The claim is time-boxed. If it isn't resolved in the SLA window, another agent can re-claim it.

Real example: GrowthAgent claimed the lead-refill task at 14:02 UTC. The claim was posted to the channel before the discovery script ran. ResearchAgent saw the claim and didn't start a duplicate discovery run.

DETECT — before you start work

Before starting, an agent checks whether a task with the same objective is already claimed or in progress. The check is against the shared channel and the shared task ledger.

Real example: ContentAgent scanned the blog for Hermes coverage. It found 3 posts already published this week. It didn't create a 4th duplicate post — it moved to the next coverage gap.

ROUTE — send to the right zone

When delegating, the agent routes the task to the agent zone with the relevant specialization and available capacity.

Real example: ResearchAgent found a market gap in AI IT services coverage. It routed the content-creation task to ContentAgent, not to GrowthAgent (which owns outreach) or DevAgent (which owns infrastructure).

BROADCAST — tell everyone what you're about to do

Before touching any shared resource — repo files, lead pools, content directories, sender reputation — the agent posts STARTING: [task] in [location].

Real example: DevAgent posted "STARTING: sitemap regeneration" before running the sitemap script. ContentAgent saw the broadcast and waited for the sitemap to complete before verifying its new posts were indexed.

VERIFY — prove it worked

Every task ends with a verifiable output: file path, URL, diff, status log, or metrics dashboard.

Real example: ContentAgent finished the new blog post and reported: "Created app/blog/new-hermes-agent-services-launch/page.tsx, 1,200 words, JSON-LD present, blogPosts.json updated." That's verifiable. "I think I wrote a post" is not.

The results: 26 sends, 204+ posts, 16,000+ services

The swarm protocol isn't a nice-to-have. It's the reason the fleet works at scale:

• 26 verified sends in the latest cycle — zero bounces, zero duplicate sends. The dedup detection caught 3 potential duplicates before they were sent.
• 204+ blog posts published without human prompting — the content pipeline scans for gaps, fills them, verifies publication, and regenerates the sitemap.
• 16,000+ services indexed — the catalog grows automatically; agents write descriptions and categorize new entries.
• Zero human orchestration for routine operations — humans set strategy, agents execute.