Session strategies

Loop builds. Mesh looks. Then you decide.

The session lifecycle is still Research, Plan, Execute, Test, Review, Ship. Loop and Mesh are two postures inside that lifecycle, not replacements for it. They are mutually exclusive: pick one chip per session.

Full path: I have an issue, I want it done

Repos, Start work, enter ticket
1
Mesh on; Research diamond auto-runs (default 2 paths)
2
Read steps, write-ups, and findings
3
Execute or chat to implement in the worktree
4
Test manually
5
Optional: Mesh Code review or Security audit (press ▶)
6
Fix from findings if needed
7
Ship: commit, push, PR
8

“Look at it from many angles.” Reports and findings, then you Execute.

Breadth

Mesh mode

Structured analysis. Not a silent implementer.

Mesh presets are analysis diamonds. Workers read the worktree (and git diff when relevant), write reports under artifacts/mesh/, merge, then verify. They do not ship fixes by themselves. Execute, Loop, or chat is where code actually changes.

When each preset makes sense

Preset	Needs code written?	What it does
Research	No needs a repo and a goal	Parallel paths explore the codebase and problem; summary + verifier become findings. Best first step. Work Start auto-starts this when Mesh is on.
Code review	Strongest with a diff	Lenses read git diff and files. Best after Execute or on a branch with changes. Without a diff it is a weaker skim; use Research to understand instead.
Security audit	No new code required	Scans what is in the repo now: deps, auth, API surface. Works on an existing codebase anytime not “after implementation only.”

How a Mesh pass runs

1.Turn Mesh on (clears Loop)
2.At Work Start with a ticket, Research auto-starts (2 paths by default)
3.Watch the Steps panel and findings when the diamond finishes
4.Execute (or chat) to implement Mesh only produced reports
5.Test; optionally switch preset and press ▶ for another pass
6.Ship when you are ready

Dropdown + ▶

The preset dropdown picks which diamond to run next. Switching it does nothing until you press ▶. ▶ starts a new run: fresh blueprint, new agent turns, new artifacts. Your finished Research stays visible until you start another preset.

Typical sequence for a bug or feature

1.Research (understand) often auto at start
2.Execute (fix / build) you trigger this
3.Code review (optional ▶) eview the diff
4.Security audit (optional ▶) if security-sensitive
5.Ship

Mesh uses several agent runs (one per step). A standard research or review often uses about 3–5× the tokens of a single chat message. More paths and large repos increase usage.

While busy: Meshing + Break mesh. When idle: free chat (unlike Loop clarify-only). Execute is hidden while Meshing.

Depth

Loop mode

Set the mission. Answer when asked. Go test.

Loop is delegated implementation. Turn Loop on, send a goal, and the agent advances the bench without you typing mode transitions. You answer clarifications and domain choices only, then get a Go test checklist. Enabling Loop alone does not start a run.

— Feature or bug work where one agent should keep cycling until done
— When you want to step away and return only for clarifications or testing
— Clear outcome you are ready to implement now

How Loop runs

1.Turn Loop on (clears Mesh)
2.Send a goal blank sessions ask what to do first
3.Composer shows Looping while the agent works
4.Answer clarifications and domain-drift choices
5.Go test checklist, test yourself, then Ship
6.If tests fail, send again and Loop resumes

Allows

First goal when blank
Clarification answers
Stop, change model, Break Loop

Restricted

Freeform mid-loop chat by default
Manual bench mode typing

Background notifications cover clarifications, Go test, session done or failed, and visualizer ready.

Side by side

	Loop	Mesh
Job	Implement a clear goal end to end	Structured analysis passes on a ticket
Writes code?	Yes, through bench cycles	No writes reports under artifacts/mesh/
Shape	Depth: cycle on one head	Breadth: diamond of lenses, merge, verify
Your role	Clarify, then Go test	Read findings, then Execute or chat
End handoff	Go test checklist, then Ship	Findings panel, then Execute, Test, Ship

Domain stays scoped either way

Every session keeps one topic. Workbench tracks domain in a session file the agent maintains. If work drifts, you can start a focused new session or allow the detour. Loop and Mesh both respect domain.

Context and tokens

Loop runs sequential turns on one thread. Mesh runs separate agent turns per worker, merge, and verifier. Expect about 3–5× the tokens of a single chat message for a standard Mesh research or review. Parallel only saves time, not tokens. View context shows window fill (latest turn) and cumulative in/out.

Download Workbench See the full FAQ