Agent workflow

How I structured my AI agent workflow

Hermes is where I started treating AI less like one chat box and more like a workflow system: profiles for different roles, memory for durable context, skills for reusable procedures, and review gates for judgment-sensitive work.

Profiles

Profiles turn AI work into assigned roles

The most useful Hermes feature to explain publicly is profiles. They make the workflow legible: different profiles are configured for different kinds of work instead of forcing every task through the same model, prompt, and context.

Input David / project need

A task arrives with project context and an intended outcome.

Coordinator Default profile

Interprets the work, selects the lane, and preserves the review boundary.

Planning + review orchestrator

Decomposition, QA, final judgment, and review gates.

UI/UX critique design

Design polish, layout critique, hierarchy, and accessibility review.

Evidence gathering knowledge-worker

Lower-cost inventory building, source clustering, and research passes.

Coding / experiments coding + local lanes

Implementation experiments, alternate hosted models, and local model serving tests.

Current working set
Profile Role in the workflow Why it matters
default Front door, context, and task routing Keeps direct requests grounded in project context and sends specialized work to the right lane.
orchestrator Planning, decomposition, review, and QA Separates judgment and acceptance checks from raw implementation.
coding Primary implementation lane Provides the strong default path for building, testing, and repairing code.
design UI/UX critique and polish Gives interface work a dedicated visual and accessibility review lane.
cybersecurity Security research and risk review Keeps security-sensitive investigation, threat analysis, and remediation advice in a dedicated lane.
knowledge-worker Research, source inventory, and documentation Uses an economical lane when collecting and organizing evidence is the job.
baseten Controlled low-cost draft lane Useful for bounded first passes; the orchestrator still gates the result before it has impact.
marketing-seo Marketing, content, and SEO work Gives audience research, content planning, and search-oriented work a dedicated lane.
inbox-intake Inbox triage and background intake Keeps recurring intake work separate from the interactive project workflow.
omlx-coder Local-model experiment lane Retains a clear place to evaluate privacy, cost, latency, and serving constraints without making it the default.

Memory + skills

Making AI work durable instead of disposable

The goal is continuity. I do not want every AI session to restart from scratch, and I do not want repeated procedures to live only in prompt history.

Memory

Durable project context

Persistent memory and Obsidian-backed notes preserve decisions, source evidence, project conventions, and review constraints across sessions. That turns AI from a one-off assistant into a continuity layer for project work.

Skills

Reusable procedures

Hermes skills act like workflow playbooks: project planning, debugging, browser automation, deployment, testing discipline, and design review. When a pattern works, it can be captured and reused instead of reinvented.

Model routing + review

The right lane depends on the job

One of the biggest lessons has been that the answer is not always “use the strongest model.” The right choice depends on cost, quality, privacy, risk, and whether the work needs reasoning at all.

01 Does it need judgment? Use a stronger model and review gate.
02 Is it evidence gathering? Use a cheaper model to collect and organize sources.
03 Is it repetitive? Use deterministic scripts instead of an LLM.
04 Does it need local constraints? Use a local model lane or guarded workflow.

Takeaway

What this taught me

Enterprise AI is workflow design. Models matter, but so do state, review, logging, access control, repeatable procedures, and knowing when not to use a model at all.