How I run my dev workflow with AI agents
My day-to-day setup: an AI agent wired into Jira, GitHub, Confluence and the browser, routed across several LLMs through an OmniRoute gateway, with a shared Obsidian vault as long-term memory, and the rules that keep it useful instead of risky.
- AI
- Workflow
- Claude
- Gemini
- OmniRoute
- Obsidian
AI didn't replace my job as an engineer. It changed what I spend my hours on. This post is a tour of the setup I use every day as a full stack developer, what it's good at, and the guardrails I put around it.
The short version
- One agent, many tools. An AI agent (Hermes Agent) runs in my terminal and editor. It talks to Jira, Confluence, GitHub and a real browser through MCP servers, so it can read a ticket, open the code, fix it and check the result in the UI.
- Many models, one endpoint. Requests go through OmniRoute, a local LLM gateway, with a priority list: Claude first, then other Claude accounts, then Gemini as the fallback. If one provider is down or rate limited, work keeps going.
- Memory that outlives the chat. An Obsidian vault is the shared brain. Every model reads it before a task and writes back decisions, root causes and conventions. It syncs to a private Git repo every few minutes.
1. The agent and its tools
A chat window that only sees what I paste in is a toy. The useful version is an agent that can act:
- Ticket in, context out. It reads the Jira ticket and acceptance criteria, then the linked Confluence spec, before touching code.
- Code with guardrails. It edits the repo, runs the linter, type-check and tests, and follows the same quality-gate rules our CI enforces (coverage on new code, no fire-and-forget promises, accessible markup).
- Verify like a human. It drives a browser with the frontend and backend running locally and walks through the acceptance criteria. The report I get is a short pass/fail list, not "it should work".
I still review every diff and make every commit myself. The agent proposes; I decide.
2. Routing across LLMs
Different models have different strengths and different bad days. I run OmniRoute as a local gateway. It exposes one OpenAI-compatible endpoint, and behind it sits a "combo": an ordered fallback chain of models and accounts:
- Claude (primary account)
- Claude (secondary account)
- A smaller Claude model
- Gemini Pro, then Gemini Flash
The agent only knows one URL and one model name; OmniRoute decides which provider answers, and its logs show which account served each request. When a token expires or a provider returns errors, the chain moves on and the session survives. The lesson: treat models like any other unreliable dependency. Add retries, fallbacks and logging, and know which account served each request.
3. Obsidian as long-term memory
Chats get compressed and forgotten; teams forget too. So every model reads one note first, Brain.md, which points to the rest:
- Learnings: my preferences, tool quirks, and how the setup works.
- Journal: one dated note per working day: what was done, decisions, open follow-ups.
- Project knowledge base: architecture, run commands, conventions, known bugs and their root causes.
Because it's plain Markdown in Git, any model can use it: Claude today, Gemini tomorrow, something else next year. The rule that keeps it clean: write the lesson, not the log. "Fire-and-forget promises need void or the Sonar gate fails" helps the next session; "fixed lint" doesn't.
4. Guardrails I don't compromise on
- No secrets in prompts or notes. Credentials live in the OS keychain or a vault; the agent fills login forms without ever seeing passwords.
- Nothing destructive by default. No pushes, no force-resets, no writes to shared databases without asking.
- Verify against reality. A model's summary is a claim, not a fact. Check the diff, the test output, the screenshot.
- Shared code stays inside approved tools. Company code only goes to providers we've agreed to use.
What changed for me
The boring parts got faster: reading tickets, chasing CI rules, writing tests for every new branch, re-checking a fix in three browsers. The interesting parts got more of my time: system design, product trade-offs, and reviewing the work properly.
This site itself was built the same way. If you're setting up something similar and want to compare notes, say hello.
Want to talk about this?
Get in touch