Ephemeral agents are contractors. Persistent agents are employees.

We tried running ephemeral agents for three months. Fresh context, clean slate, no baggage. It felt pure — like hiring a brilliant contractor who shows up, does perfect work, and leaves.

The reality was different. Every conversation started with 20 minutes of context rebuilding. "Here's our codebase structure." "This is how we deploy." "Remember, we use this specific API pattern." By the time the agent understood our setup, half the session was burned.

Then we switched to persistent agents. Same agent, same memory, continuous context. It felt messy — like having an employee who remembers every weird decision you made six months ago.

But here's what we learned: the contractor model works for discrete tasks. The employee model works for ongoing operations.

Our ephemeral coding agent was perfect for one-off scripts:

Task: "Build a CSV parser for user imports"
Result: Clean, focused, no assumptions
Time to value: 15 minutes

But for ongoing development work, it was brutal:

Session 1: "Here's our auth system..."
Session 2: "Here's our auth system again..."
Session 3: "No, we talked about this yesterday..."

Our persistent agent remembers our deployment pipeline, knows our coding style, and can pick up where it left off. But it also remembers that time we tried a terrible architecture pattern and argued about it for an hour.

The hybrid approach: Use ephemeral agents for discrete deliverables. Use persistent agents for ongoing partnerships.

We now run ephemeral agents for:

• One-off scripts and utilities
• Code reviews (fresh eyes, no bias)
• Research tasks with clear endpoints
• Anything where "clean slate" is a feature

We run persistent agents for:

• Feature development across multiple sessions
• Bug investigation and fixes
• Architecture decisions that span days
• Anything where "knows our context" matters more than "fresh perspective"

The key insight: ephemeral agents are contractors, persistent agents are employees. Hire contractors for projects. Hire employees for processes.

Our persistent coding agent now carries forward context about our codebase, our patterns, and our preferences. It doesn't waste time relearning our auth system every session. But when we need a fresh perspective on a gnarly bug, we spin up an ephemeral agent with just the relevant code.

The cost difference is dramatic. Ephemeral agents burn tokens on context rebuilding. Persistent agents burn tokens on memory maintenance. For ongoing work, persistent wins by 60%. For one-off tasks, ephemeral wins by 40%.

Most teams pick one model and force everything through it. The teams that win use both strategically.