Big work

Hand big work to a swarm.

For big tasks and whole projects: Toad spawns mini-bosses that decompose the goal, orchestrate a swarm of agents, self-correct, and keep going through phase gates until the plan is actually done — so you get completion visibility instead of driving every pane. The pod view, Lilly Pod, shows the swarm breathing.

At a glance

Reach for a swarm for big tasks and whole projects — not everyday delegation.
Mini-bosses decompose the goal, orchestrate agents, self-correct, and push through phase gates until done.
A swarm is scoped, not infinite — each mini-boss runs under a per-boss child cap.
The pod view (Lilly Pod) shows the swarm breathing: active bosses, children, success rate, and alerts.

Note

This is for big work. Most delegated work goes to background agents and task groups — that is the default, no swarm required. Reach for a swarm only when the goal is genuinely big: multi-phase lifecycle, large fan-out, or overnight and unattended runs. A swarm is scoped, not infinite — each mini-boss runs under a per-boss child cap and moves through phase gates, so it stays steerable.

The pod view turns background work into a readable topology: active bosses, child agents, success rate, alerts, and done counts — a whole swarm at a glance.

lilly pod — live swarms

Autonomous swarms with live counts and alerts.

Describe the big goal

Give Toad the outcome, not each terminal command. State what done looks like — the mini-boss handles the plan and the breakdown. Toad decides how to route the request, so a genuinely big, multi-phase goal lands as a swarm rather than a single background agent.

Mini-bosses orchestrate the swarm (per-boss child cap)

Toad stands up mini-bosses that split the goal and spawn a swarm of child agents, assigning each one a slice of the work. Every mini-boss runs its children under a per-boss child cap — the swarm is scoped, not infinite, so it stays steerable and never spirals. The brain can run on almost any model, but the swarm's agents do the coding; the Gemini and Antigravity CLIs are agents-only and never drive the brain.

Self-correction runs automatically

When a child stalls or fails, the boss reruns and repairs it on its own — you do not have to babysit each step. Self-correction is part of the loop, so the swarm keeps making progress without you stepping in for every hiccup.

Phase gates keep it on track

The swarm advances through phase gates: each phase has to clear before the next begins, and the pod view shows children, progress, and alerts in real time. Status badges flag any swarm that needs a decision, so you can steer at the gate instead of watching every agent. Work that pulls in a repository reaches GitHub through the gh CLI.

Get completion visibility (each swarm stays tied to its workspace)

The boss keeps going until the plan is actually done, then reports back. Each swarm stays tied to its workspace, so outputs and follow-ups land in the right repo context — across workspaces, from one pond. If a swarm reaches out to a confined remote agent, that work stays approval-gated; nothing lands without your say-so. Want to watch a slice live instead of at the gate? Spawn an Agent View pane and drive it by hand.

When to reach for it

Reach for a Swarm (Lilly Pod) when the job is too big to watch and too long to drive by hand — a large task or a whole project.

Multi-phase goalsWork that moves through several stages a boss can sequence and coordinate through phase gates.

Large fan-outMany parallel slices the boss spreads across child agents at once, under the per-boss cap.

Overnight runsUnattended work you start and check on later instead of watching live.

Completion visibilityYou want the result and a report, not to steer every step yourself.