Why inline agent?

Not all of the coding ends up in agentic-coding

They are

expensive
slow; long turnaround time
inconsistent w.r.t. UX
- editor ←> terminal
- writes can overlap
- context switch

On the other hand, Editor inline agents are

consistent w.r.t. UX
- everything inside editor
well-informed of your problem and code
- can access your code & doc

Additionally, as many hand-coding problems are “simpler” ⇒ faster models are ideal in this case, such that:

cheaper
faster

Workflow

code with inline agent

flowchart LR
  subgraph editor
    direction LR
    code --> try_run
    try_run -->|runs OK| code
    try_run -->|fails| ask
    ask --> code
  end

Specifically for Zed editor:

cmd + r to open up agent panel
cmd + n to start new session
set agent.tool_permissions.default = "allow"

Conventional coding (baseline)

flowchart LR
  subgraph editor
  code --> try_run
  try_run -->|runs OK| code
  end
  subgraph browser
  try_run -->|fails| search
  search --> code
  end

The “browser” part is not efficient:

UX perspective
- different application for a single problem
Context switch
- Browser does not know your context
- You need to compress your problem in short