Aaron Ahmad

Hermes Agent: Multiple Profiles

07 August 2026 · tech
Patent-style schematic of multi-profile agent architecture

It seems like everyone is talking about a Chief-of-Staff. An AI agent that will be your right hand throughout your day. You're Washington. It's Hamilton.

I've been running a Hermes Agent for the last few months and I am 100% in on this idea. Whilst I'm definitely having to bend myself/my tasks in a way to make the Agent able to fit, there is definitely value in the work I have it do for me. This website for one. I would never had had the time or patience to build something even this basic. With my Agent I just iterate and direct. One annoyance with the Agent was that it was always busy doing the tasks I assigned. A bit of digging and back and forth I realised this is a solved problem and there was a better way; boards and profiles.

FIG. 1 SHEET 1/3 MULTI-PROFILE SCALE N/T ARCHITECTURE USER OPERATOR REQUEST CONDUCTOR ROUTING & DISPATCH ISOLATION BOUNDARY SPECIALIST A WEBSITE MODEL: LIGHT SPECIALIST B CODE MODEL: HEAVY SPECIALIST C RESEARCH MODEL: REASON SPECIALIST D VAULT KEEPER MODEL: LIGHT SHARED ACCESS TASK BOARD SHARED QUEUE COORDINATION VAULT SHARED STORAGE KNOWLEDGE BASE
FIG. 1 — Multi-profile architecture schema: conductor dispatches tasks to isolated specialists, all sharing a task board and vault.

Instead of one AI assistant doing everything in a single chat window, I have a team of isolated agents - each with its own context, model, and permissions. A Conductor routes every request. Specialists handle the work: one keeps the website running, another writes code, a third does research, another maintains my knowledge vault. They coordinate through a shared task board and a shared documents folder. This isn't like the incessant YouTubers talking about how they have 100 agents running 24/7 or have built some neon html dashboard to chat to each agent. That shit is shit and is a waste of time and just bait.

FIG. 2 SHEET 2/3 TASK FLOW SCALE N/T TASK LIFECYCLE — SEQUENTIAL FLOW DIAGRAM ① OPERATOR SUBMITS REQUEST TO CONDUCTOR ② CONDUCTOR ROUTES TO MATCHING SPECIALIST ③ SPECIALIST READS FROM / WRITES TO VAULT ④ SPECIALIST LOGS COMPLETION ON TASK BOARD USER STEP 1 REQUEST ① CONDUCTOR STEP 2 ROUTE TASK ② SPECIALIST STEP 3 EXECUTE ③ READ VAULT SHARED KNOWLEDGE ④ UPDATE TASK BOARD STEP 4 LOG RESULT ⑤ RESULT STEP 5 DELIVER ⑤ RESULT DELIVERED TO OPERATOR
FIG. 2 — Task lifecycle: a request flows from user through the conductor to a specialist, which reads the vault, logs to the board, and returns the result.

The setup allows me to converse with a single agent. When I need something done that I know is multiple actions, or is for a specific project, I give my agent to goals or tasks I think we need done and it posts them to a board for other agents to action and report back. The setup helps because context stays clean. The agent managing my website doesn't carry my research notes. The agent writing code doesn't have access to my database. Each profile uses the model that fits the job - cheap and fast for simple work, heavy reasoning only when needed.

FIG. 3 SHEET 3/3 SHARED STATE SCALE N/T SHARED RESOURCES — VAULT & TASK BOARD ACCESS MAP PERSISTENT KNOWLEDGE BASE LIVE COORDINATION QUEUE VAULT SHARED STORAGE NOTES · DOCS · MEMORY TASK BOARD SHARED QUEUE COORDINATION PROF. A PROF. B PROF. C PROF. D EACH PROFILE INDEPENDENTLY READS & WRITES TO BOTH SHARED RESOURCES LEGEND: VAULT ACCESS BOARD ACCESS
FIG. 3 — Shared resources map: every profile independently accesses the vault (persistent knowledge) and the task board (live coordination queue).

What I've learned so far is isolation is a strength. Keeping tasks or skills separate has helped me create agents that can do one thing well and also helped my understanding of whats going on and, more often than not, why something went wrong. I can ensure one agent doesn't have access to a folder so I know it cant get in there to run amok. Separate permissions mean smaller blast radius. What effect this has on token usage, I dont know. Something for me to try and track. But for now the coordination overhead is worth it.

← All posts Reply via email