Loop builds. Mesh looks. Then you decide.
Every session follows the same steps: Research, Innovate, Plan, Execute, Review, then test and ship. Loop and Mesh are two ways to run it: keep building until it's ready to test, or look from many angles first. Pick one per session.
From ticket to done
Open the repo and start with your ticket
1Turn Mesh on. It explores the problem from a few angles
2Read the write-ups and the findings
3Build the fix yourself or let the agent implement
4Test the change
5Optional: another Mesh pass for review or security
6Ship when it looks right
7
“Look at it from many angles.” Findings first, then you choose what to build.
Mesh mode
Structured looks. Not a silent implementer.
Mesh is for understanding and critique, not silent shipping. It explores the repo from a few angles, writes findings you can read, and stops there. You choose when to build: Execute, Loop, or a normal chat.
| Preset | Needs a change first? | What you get |
|---|---|---|
| Research | No, just a repo and a goal | Several paths explore the codebase and the problem, then you get a clear summary. Best first step. Often starts automatically when Mesh is on at work start. |
| Code review | Strongest with a change already made | Looks at the diff and related files. Best after you have built something. Without a change, prefer Research to understand first. |
| Security audit | No new code required | Checks what is in the repo now: dependencies, sign-in and access, and your API. Works on any existing codebase, not only after a feature lands. |
- 1.Turn Mesh on
- 2.At work start with a ticket, research often starts for you
- 3.Watch the steps and open the findings when they land
- 4.Build it yourself or with the agent. Mesh only does the looking
- 5.Test; run another Mesh pass if you want a second lens
- 6.Ship when you are ready
Choose the preset for the next pass. Changing the dropdown does nothing until you run it. Each run is a fresh look; finished research stays visible until you start another preset.
- 1.Research: understand the problem (often at the start)
- 2.Execute: build the fix
- 3.Code review: an optional second look at the changes
- 4.Security audit: optional for sensitive work
- 5.Ship
Mesh uses several agent runs. A standard research or review often costs about a few times a single chat message. More paths and larger repos mean more spend.
While Mesh is running, the status shows Meshing and you can stop it. When it's idle, chat freely. Execute waits until the Mesh pass is done.
Loop mode
Set the mission. Answer when asked. Go test.
Loop is “keep building until I need you.” Turn it on, send a goal, and the agent drives the session. You mostly answer its questions, then you get a checklist to test. Turning Loop on doesn't start a run by itself; you still send.
- A feature or bug where one agent should keep going until it is done
- When you want to step away and only come back for questions or testing
- A clear outcome you are ready to implement now
- 1.Turn Loop on
- 2.Send a goal (blank sessions ask what to do first)
- 3.Watch while it works; answer when it asks
- 4.Get a test checklist, try it yourself, then ship
- 5.If something fails, send again and Loop continues
You can
- Your first goal
- Answers to the agent's questions
- Stop, change model, or break Loop
Not the default
- Free chat in the middle by default
- Hand-driving every phase yourself
When Workbench is in the background, you get notified for questions, ready-to-test, failures, and visualizer ready.
| Loop | Mesh | |
|---|---|---|
| Best when | You have a clear goal and want it done | You need to understand before you change code |
| Writes code? | Yes, the agent builds as it goes | No, it gives you findings to act on |
| Shape of the work | Depth on one goal, step after step | Breadth: several angles, then one clear summary |
| Your job | Answer questions, then test the result | Read the findings, then decide what to build |
| When you are done | A checklist to test, then ship | A findings report, then build, test, and ship |
Stay on one topic either way
Each session keeps one focus. If the work wanders, the agent stops and asks: start a fresh session, allow it once, or allow it for the rest of the session. Loop and Mesh both respect that.
What it costs to run
Loop is one thread of work, step after step. Mesh runs several looks at once, so a research or review pass usually costs a few times more than a single message. Running in parallel saves time, not money. Open View context to see how full the session is getting.