Architecture Decision: Why career-ops uses a single persistent agent

[santifer]

A recent PR (#243 by @tharun-a-sure) sparked an important conversation I want to share with the community. The PR is a beautifully designed web dashboard — 10 pages, chat integration, Catppuccin Mocha theme, vanilla JS with no build step. Seriously impressive work.

But reviewing it made me articulate something I've been thinking about for a while: why career-ops is built around a single persistent agent, and why that matters.

The core insight

career-ops serves one person per instance. It's not multi-tenant. It's not a SaaS. It's your personal career advisor running on your machine.

This changes everything about how it should be architected.

Two approaches to AI agents

Spawn-and-kill (ephemeral agents):

Each command launches a new AI process. It responds. It dies. Next command, new process.

• You evaluate a job at Anthropic → agent scores it B+
• You ask "how does this compare to the Stripe role I saw yesterday?" → agent has no idea what you're talking about
• You generate a CV → agent doesn't know that your last 5 evaluations all flagged "leadership experience" as a gap
• Every interaction starts from zero — like going to a different doctor every visit

This is how most AI automations work today (n8n, Make, Zapier). Stateless pipelines. Great for data processing. Terrible for career advising.

Single persistent agent (career-ops approach):

One agent session. Context accumulates. The agent gets smarter about you over time.

• You evaluate 40+ jobs over a week → the agent notices patterns: "you keep scoring infrastructure roles higher than frontend"
• It learns your preferences → "you rejected the last 3 roles under $150K — adjusting thresholds"
• Your CV improves with each iteration → the agent remembers what hiring managers responded to
• It connects dots across your entire pipeline → "this company rejected you 6 months ago but just posted a senior role — different team, worth re-evaluating"

Like having a career advisor who knows your entire history.

Where the industry is going

This isn't just our opinion. The entire industry is moving toward persistent agents with memory:

• Claude Code has memory files that persist across sessions
• ChatGPT introduced cross-conversation memory
• Cursor maintains full project context
• Every major AI lab is investing in long-context and agent memory

The value isn't in running isolated commands with a pretty UI. The value is in depth of context — an agent that understands your career trajectory, not just your latest job description.

What this means for our roadmap

The desktop app on our roadmap (🟣 Later phase) will NOT be spawn-and-kill. It will be:

• A visual interface on top of a persistent agent — not a replacement for the brain
• One agent, living memory, beautiful UI
• Think VS Code wrapping a language server — the UI enhances the intelligence, it doesn't replace it

For contributors

This doesn't mean we don't want UI work — we absolutely do. But any UI contribution should:

1. Preserve agent persistence — don't spawn independent processes per command
2. Build on the existing context system — pipeline, tracker, evaluation history, profile
3. Enhance the conversation, not fragment it into isolated interactions
4. Respect security — no bypassPermissions, no unauthenticated endpoints that execute commands

PR #243 is great visual inspiration for what the future desktop app could look like. The design work is genuinely impressive and we'll reference it when we get there. The architecture just needs to match the philosophy.

Join the conversation

We just opened a #devs channel in Discord for exactly these kinds of architectural discussions. If you're interested in building the desktop app, working on the agent persistence layer, or have thoughts on stateful vs stateless agents — come talk to us.

Let's build together.