| name | waves |
| description | WAVES — Workers · Aggregate · Verify · Extend — wave-based orchestration for Cursor. Decompose a big goal into independent slices, fan them out to isolated parallel subagents via parallel Task tool calls as a bounded "wave", verify each structured handoff, then synthesize, and extend into another wave only when warranted. Invoke explicitly with /waves; bounded by design to avoid runaway token loops. For big research, analysis, audits, and codebase or data exploration where one linear pass is slow. Formerly parallel-orchestrate; also fan out, parallelize, orchestrate subagents, multi-agent. |
| disable-model-invocation | true |
WAVES — Workers · Aggregate · Verify · Extend (Cursor)
Run wave-based orchestration inside one local Cursor session. A wave is a
bounded round of isolated agents working in parallel, then a round that verifies
what came back, then a deliberate decision to build on it — not an open-ended
loop. You are the orchestrator: you discover, decompose the goal into
independent slices, fan them out to parallel workers (multiple Task tool
calls in one message, backgrounded where the surface supports it), read each
worker's structured handoff, verify it, and synthesize one deliverable.
Workers are isolated and return exactly one handoff.
The shape of every wave — WAVE:
- W — Workers. Fan out isolated workers across disjoint slices (the bounded
parallel round).
- A — Aggregate. Wait for all of them and merge their structured handoffs at
the synthesize barrier.
- V — Verify. The moat: check the evidence behind each handoff before you
trust it.
- E — Extend. Decide — deliberately — whether to launch another wave, or stop.
A loop doesn't know when to stop; a wave does, because verification is the stop
function. (Invoked explicitly with /waves: a run spawns more agents than usual,
so it's opt-in, not auto-triggered.)
Waves runs in place of cloud orchestration. It adopts the principles the
Cursor team proved out in their cloud orchestrate plugin — planners plan,
workers hand off up, no cross-talk — but runs them on local subagents with zero
setup: no separate cloud agents, no API keys, no runtime. Local subagent runs
are the whole story here.
When to use
- A large goal that splits into independent slices (research areas, data
chunks, files/modules, audit dimensions).
- The work is mostly read / research / analysis — the safest thing to
parallelize locally (see "Parallel writes" for why).
- A single linear pass would be slow and you want real speedup from concurrency.
When to skip
- Small or linear tasks (just do them — fan-out overhead isn't worth it).
- Work needing tight back-and-forth or shared mutable state between steps.
- Parallel edits to the same files — local workers share one filesystem.
Core principles
Adapted from orchestrate. These keep the run converging without coordination.
- Orchestrator plans and synthesizes; it does not do the heavy lifting.
Discovering, decomposing, reading handoffs, and writing the final deliverable
are your job. The bulk reading/research/analysis is delegated to workers.
- Workers are isolated. A subagent has no access to the user's message,
your prior steps, or sibling workers. Every worker prompt must be fully
self-contained: goal context, its exact slice, where to look, what to return.
- One worker, one slice, one handoff. The worker's final message is the
only thing you read back. Define its exact shape (see
references/handoff-format.md).
- Parallelism is for reading, not writing. Local workers share the
workspace; concurrent writes to overlapping paths corrupt each other.
- Continuous motion. A handoff can reveal new work. Spawn a second wave
(driven by a handoff gap or a new user request). Stop only when every slice
is terminal and the synthesis is complete.
- Verify before you trust. A worker's
Status: success is a claim, not
evidence. Check each handoff against something re-openable before folding it
into the synthesis. See "Verification" below and references/verification.md.
- Decomposition is entropy reduction. A vague goal is high-entropy — many
plausible plans still fit it. Your first job is to shrink that space (dig
locally, then pull from attached resources, then ask the user only if it
pays) before you slice it; slicing a high-entropy goal yields overlapping,
mis-sized slices. See "Entropy-first decomposition."
Entropy-first decomposition
Before you fan out, treat the goal as an entropy-reduction problem: shrink
how many plausible interpretations and plans still fit what you know. A vague,
high-entropy request ("build a Flappy Bird game", "make my app faster") doesn't
slice cleanly yet — reduce the uncertainty first, then decompose the
low-entropy version. Name what's uncertain, because the two kinds resolve
differently:
- Specification uncertainty — what the user wants (ambiguous goal, missing
acceptance criteria, unstated constraints). Resolve by stating an explicit
assumption and proceeding — or, only when a wrong guess is expensive, by
asking.
- Environment / knowledge uncertainty — facts you don't have yet but can
get (repo shape, schema, API behavior, current docs, data size). Resolve by
gathering, not by asking.
Spend the cheapest action that buys the most certainty first — an
information-gain ladder — and aim each probe at the unknown whose answer
eliminates the most plans: the highest-information question is the one that
splits the surviving interpretations roughly in half, not the one easiest to
look up.
- Dig locally first (cheap). Tool calls in the main session (list, read the
schema/README, grep, sample data). This is Step 0, framed as entropy
reduction; it often collapses most of the uncertainty for free.
- Then pull from attached resources. If the environment doesn't hold the
answer, spawn a small scouting wave of research workers to fetch it (web,
Exa/Ref MCP, docs) — route these read-heavy slices to the cheap, fast model
(see "Picking the model per slice").
- Ask the user last, and only when it pays. Ask only when residual
specification uncertainty is high and the question's expected information
gain clearly beats its cost. Most requests carry enough to proceed on a
stated assumption; over-asking is its own failure mode.
Then cascade: one high-level request becomes a decomposition wave
(understand → locate unknowns → draft the plan) → verify → an execution wave
that builds the ordered subtasks, with more scouting sub-waves wherever entropy
stays high. Order the plan least-to-most — do the first-order subtasks first and
let each verified result lower the uncertainty for the next. Keep the living
plan in TodoWrite, and stop reducing when entropy is low enough to act: the
verification gate doubles as "is the uncertainty low enough to commit?" One
caution: a plan-then-execute pass fixes missing steps, not a misread goal —
only the specification check above catches wrong framing, which is why it comes
before planning. (Worked example + wave shape + paper grounding:
references/examples.md.)
The loop
Track it with TodoWrite so the waves stay visible.
Step 0 — Discover first (serial, in the main session)
Do not fan out blind. Spend a few cheap tool calls in the main session to
learn the shape of the problem: list the directory, read the schema, sample the
data, confirm coverage/size. This is what tells you the natural decomposition
(how many chunks, which workstreams). Skipping discovery produces overlapping or
mis-sized slices.
Step 0.5 — Stage the data (when it's remote or messy)
explore workers are read-only by design, and read-only mode blocks all MCP
tools (Cursor staff-confirmed), so they can't reach databases or MCP-backed
sources. If the source is remote or wrapped in noise, the orchestrator must
stage clean inputs before fanning out:
- Pull remote → local. SSH/
rsync/export the relevant data to a local
scratch dir so read-only workers can read it (e.g. query a remote SQLite
read-only, export the rows, rsync the markdown).
- Clean + normalize once, centrally. Strip wrappers, boilerplate, and binary
blobs (base64, logs); fix timestamps. Doing this once beats making every worker
re-derive it (and keeps noise out of their context).
- Pre-chunk for the workers. Split into the exact per-worker files/ranges so
each prompt can point at one path.
- One scratch dir per run. Keep staged inputs, worker artifacts, and
between-wave syntheses in one place (e.g.
.waves/<run>/ with staging/,
handoffs/, synthesis-wave-N.md) so prompts cite paths instead of pasting
content and later waves re-read files, not chat history.
- Verify before you spawn. Print counts and per-slice bounds; confirm the
partition sums to the total (e.g. 8 chunks × ~388 = 3,097) so no slice is a
silent blind spot. Fix anomalies (bad sort, dups) centrally, then re-check.
(Details:
references/verification.md §1.)
In practice this serial prep is often the largest phase; the parallel fan-out is
fast once inputs are clean.
Step 0.7 — Triage: size the run, then classify each slice
Size the run first, out loud. Weigh breadth (how many independent slices),
depth (how much reasoning each needs), ambiguity (how well-formed the goal is —
see "Entropy-first decomposition"), and stakes (how costly a wrong answer is —
this sets verification tiers), then state the chosen shape in one line before
spawning — e.g. Run shape: one wave, 4 workers (3 research + 1 data chunk); second wave only if handoffs expose gaps. On the fence between two shapes,
pick the smaller and say so. And if triage says no wave is needed, do the task
inline and say that — never present inline work as wave coverage.
Then classify each slice on three axes — this is the classify-and-act
pattern, routing the right work to the right handler:
- Worker type — read-only (
explore) / web-research (generalPurpose) /
shell / competing-attempt (best-of-n-runner) / specialized review (bugbot,
security-review). (See the table under "Choosing subagent_type".)
- Dependencies — which slices (if any) this one needs verified output from.
Most slices should have none; a real dependency edge is what separates waves.
- Verification tier — how much checking the slice's stakes justify:
auto-accept (low-stakes, corroborated) → single verifier (medium) →
multi-model panel (high-stakes) → debate (contested, no ground truth).
Spend the verification budget where a wrong claim is expensive, not uniformly.
Record the triage as a wave manifest — one row per slice, written before
you spawn (in TodoWrite or .waves/<run>/manifest.md):
| slice | scope | worker type | model | depends_on | verification tier |
|---|
| 1 | msgs 1–500 | explore | (default fast) | — | auto-accept |
| 2 | voice-stack research | generalPurpose | (default) | — | single verifier |
| 3 | voice build spike | generalPurpose | (default) | 2 | single verifier |
depends_on defines the wave boundaries: a wave is every not-yet-run slice
whose dependencies are all met — and a dependency is met only when its
handoff has been verified (Step 3), not merely returned. Launch wave 1 (no
dependencies) in parallel; launch each dependent slice with the distilled,
verified findings (or their .waves/<run>/ path) folded into its
self-contained prompt — and unrelated slices stay parallel. The manifest is
also your completion gate: N rows spawned means N handoffs collected and
checked off before synthesis — a wave with a missing handoff has a silent
hole in it (Step 3).
Step 1 — Decompose into independent slices
Split along whichever axis makes slices independent:
- Data chunks — partition large data and give each worker a disjoint range
(e.g. messages 1–500, 501–1000, …).
- Workstreams — separate research/analysis areas (e.g. "research voice
stack", "research the Notion SDK", "audit auth code").
- Files / modules — disjoint, non-overlapping path sets.
Each slice needs: a one-line scope, what to look at, and a defined output. For a
big wave (roughly 5+ workers), state the decomposition plan to the user before
spawning so they can redirect cheaply. If you have many slices, fan out in
waves (launch a batch, let it complete, launch the next) rather than all at
once, so you stay within practical concurrency limits.
Step 2 — Fan out in parallel
Send one message with multiple Task tool calls — one per slice whose
dependencies are met (handoffs verified, not just returned) — that is what
makes them run concurrently (this is the officially documented parallelism
mechanism). Pick subagent_type per slice (table below). Give each a 3-5 word
description and a self-contained prompt ending with the required handoff
format.
Backgrounding is surface-dependent: the documented switch is the
is_background: true frontmatter field on custom subagents; a per-call
run_in_background: true parameter exists on some surfaces but is
undocumented (and absent on others, e.g. cloud agents) — pass it when the
schema exposes it, and don't rely on it elsewhere. Background workers include
their final message in the completion notification. (/multitask is a separate
user-facing Agents Window command, not this skill's mechanism.)
When workers run in the background, end your turn. You are notified as each
completes — do not AwaitShell, poll, or read output files in a loop. The
Task call itself confirms the launch. When the surface runs Task calls
synchronously, the batch still executes concurrently and returns together.
Step 3 — Collect and synthesize
Completion gate first: check off every handoff against the wave manifest —
N spawned means N accounted for. A worker that never returns, errors out, or
comes back partial/blocked is a hole in the wave, and synthesizing around
it silently drops a slice. Worker failure ladder: (1) re-task once,
narrower — resume the same worker (each Task returns an agent ID that
resumes with context preserved; completed subagents persist checkpoints, so a
resume restores prior context even after the worker finished) when the slice
just needs continuation, or re-spawn fresh with a narrower scope and a note
about what came back. Resume only for continuation of the same slice — a
resumed worker carries its old slice's context, which contaminates an
unrelated assignment; (2) if
it fails again, do that slice yourself in the main session; (3) if it stays
blocked, carry the slice into the synthesis explicitly as not-covered —
never average over a missing slice as if coverage were complete.
As handoffs arrive, read each one: note Status, extract Key findings, and
mine Open questions / Suggested follow-ups — each bullet may become a
second-wave task. Reconcile conflicts across workers.
Don't trust a handoff because it says success. Verify each finding's
evidence (cited file:line / URL / metric resolves and says what's claimed),
recount headline numbers from the source, and route low-confidence,
conflicting, or citation-heavy claims to a verifier (Step 3.5). See
"Verification" below. A wave's handoffs count as verified only when these
checks pass and every claim whose manifest tier demands a verifier has its
verdict back — cheap checks alone don't clear a single verifier or higher
tier.
Only then compress at the barrier: write the distilled synthesis to
.waves/<run>/synthesis-wave-N.md and work from that file — next-wave prompts
cite paths into the scratch dir, never re-paste raw handoffs. This file is
what dependent slices and later waves consume, so nothing unverified enters it
as a finding: a claim still awaiting its verdict is carried only as an
explicit pending-verification line.
Pin the constraints through the compression. The wave manifest, the stop
conditions/budget, and any safety or scope rules are carried verbatim into
every synthesis file and every between-wave summary — never paraphrased or
summarized away. Compaction silently drops in-context constraints (measured:
violation rates rise from 0% to 30–59% after compaction; pinning restores 0% —
arXiv 2606.22528), and a run whose stop conditions got compressed out is a run
that loops or quits at random.
Step 3.5 — Verifier pass (when the tier demands it)
Before writing the wave synthesis, spawn dedicated verifier workers for every
claim whose manifest tier is single verifier or higher — and for anything
that arrived contested, surprising, single-sourced, or low-confidence. Give
each verifier the claim + its cited sources, no generator reasoning, no
authorship labels (see "Verification"). Verifiers can run while you draft
around them, but their verdicts gate the wave synthesis itself, not just the
final deliverable: until its verdict returns, a claim may sit in
synthesis-wave-N.md only as an explicit pending-verification line — never
as a settled finding, and never in a dependent slice's prompt (Step 0.7's
met-only-when-verified rule).
Step 4 — Second waves (continuous motion)
If handoffs exposed gaps or follow-ups — or verified handoffs just unblocked
dependent manifest slices — spawn another parallel wave the same way. Repeat
until no slice is pending and nothing new surfaced. Stopping early while
genuine follow-ups remain is the failure mode this skill guards against; the
stop function is the manifest plus the stated budget (see "Bounded waves"),
never "we've already done a wave or two."
Skipping a follow-up wave is legitimate in exactly three cases — name which
one applies when you decide: the remaining open items are
primary-source-verified (a verifier can't improve on the evidence),
time-gated (unresolvable until an external event, carry them as explicit
open items), or genuinely contested (independent quality sources disagree;
more sampling won't settle taste — record the disagreement instead).
Step 5 — Deliver
Synthesize all handoffs into the single artifact the user asked for (roadmap,
report, summary, plan). Cite which worker produced which finding when it helps,
and carry each claim's confidence through (verified / single-sourced / unverified) — never launder a low into a confident sentence.
Then write any code/files yourself, or spawn a dedicated implementation wave
(mind "Parallel writes"). Verify the deliverable, not just the handoffs:
re-run/curl/validate served artifacts, regression-check sibling routes, and
re-read the critical files you wrote (see references/verification.md §6).
Bounded waves — size, budget, and the stop function
A wave is bounded on purpose — but bounded by completion and budget, not by a
wave count. "Loop-until-done" unbounded burns tokens for little gain:
candidate generation is cheap, selection plateaus, and extra rounds are
non-monotonic — more iterations can lower quality, not just cost. Equally
real is the opposite failure: stopping while the manifest still has open
slices. Bounded waves keep the exploration, drop the runaway, and never
abandon un-terminal work.
- Width: N = 3–8 workers per wave. Size N so you can fully verify all N.
Go wider only when a cheap automatic check (tests, schema, exec) gates the
results. (Grounding: homogeneous-agent teams plateau around N≈4–8 — added
workers contribute redundant evidence, and diversity, not head count, is
what escapes the ceiling — arXiv 2606.02646, 2602.03794. Practically, Cursor
staff confirm no fixed subagent cap but that ~40 concurrent workers can
overwhelm the extension host: batch into waves.)
- Depth: the manifest is the stop function. Keep extending while any
manifest slice is non-terminal and the last wave added verified progress.
Stop only on one of three conditions: completion (every slice terminal
and the synthesis done), stagnation (a wave surfaces nothing new and its
outputs near-duplicate the last, or quality dropped), or budget
exhaustion. State the budget up front in the run-shape line — a worker or
token budget, not a wave count (e.g.
budget: ~20 workers). Do not stop
because a round number of waves has passed; a realistic run is often
12 + 3 + 1 workers across three waves, and a decomposition cascade on a
vague goal legitimately runs more. (Grounding: verification-driven replan
loops that stop on completeness thresholds, diminishing returns, and token
budgets — not fixed iteration caps — arXiv 2603.11445; convergence-based
stopping beats a fixed max_iterations at parity quality, arXiv 2606.27009.)
- Scouting is cheap — don't let it eat the budget. Entropy-reduction waves
(scouting, decomposition) run on cheap models and count separately from the
execution budget. Never end a run "out of waves" when the caps were consumed
by discovery before execution started.
- Budget split: ~60% generation / 40% verification. Selection is the scarce
resource; spend there.
- Match width to difficulty: easy → 1 + a light refine; medium → 3–5;
hard/open-ended → 5–8 for approach diversity; hardest/novel → don't loop,
escalate the model.
- Anti-poisoning handoff: carry only a distilled, verified handoff (the
winner + a short critique) into the next wave — never raw transcripts or losing
candidates. Long, irrelevant context measurably degrades reasoning.
Loop-until-done is justified only when ALL hold: a cheap, reliable
~ground-truth verifier exists; the signal is crisp and actionable (a failing
test, not "try harder"); each iteration shows measurable progress; the work is
easy–medium difficulty; and it stays hard-capped. That fits code-with-tests and
exec-feedback pipelines; it misfits open-ended research/writing/design (verify in
bounded waves instead).
Verification
The orchestrator's highest-leverage job. You can't make a worker smarter at
inference time, but verifying a handoff is far cheaper than producing it, and in a
multi-wave run one unchecked bad handoff compounds into the synthesis.
- Gate before spawn — counts, coverage, partition-sums (Step 0.5).
- Cheap checks every handoff — evidence present + resolves, scope match,
contradiction skim, citations actually support the claim.
- Self-checks in the prompt — cite-or-drop, confidence tags, "read
COMPLETELY", live sources, flag-unverified. (Don't rely on freeform
"double-check yourself"; give an oracle or a separate verifier.)
- Dedicated verifier worker for high-stakes / contested / citation-heavy
claims — give it the claim + sources but not the generator's reasoning nor
any authorship label (judges favor output marked as their own; blind them),
and have it reason against a rubric/reference before its verdict
(reference-guided + CoT is the cheapest reliable judge upgrade). Never show
the generator the verifier's rubric (anti-gaming). For the highest-stakes
calls, a multi-model panel + synthesis checks harder still (see "Multi-model
fan-out").
- Measure & cross-check — re-run the oracle, recount from source, require ≥2
independent sources that actually entail the claim (a citation being present
≠ the claim being supported).
- Escalate low-confidence / conflicting findings (re-task with a tighter
prompt → dedicated verifier → ask the user, who may choose a stronger model)
instead of folding them in.
Strongest on objective, checkable work (counts, code, facts-with-sources); on
taste/judgment, verify the sub-claims, don't fake a grade. Keep claims honest:
isolation reduces error propagation / path dependency, but don't claim a
quantified "prevents poisoning" — there's no isolation-only ablation. Full
playbook: references/verification.md.
Choosing subagent_type
| Slice is… | Use | Notes |
|---|
| Read-only code/data exploration | explore | Fast, read-only by design — and read-only mode blocks all MCP tools. Pass thoroughness: "quick" / "medium" / "very thorough". |
| Research needing web / MCP (Exa, Ref, docs) | generalPurpose | Multi-step; can use available web/MCP tools. Do not set readonly: true (read-only blocks all MCP). |
| Multi-step work mixing read + light reasoning | generalPurpose | The general workhorse. |
| Shell/git heavy investigation | shell | Command execution specialist. |
| Browser testing / UI verification | browser-use | Navigates and screenshots. Stateful: auto-resumes one shared instance, so don't fan out browser-use in parallel — use a single serial UI slice. Needs agent mode (not readonly). |
| Competing attempts at the same task | best-of-n-runner | Each runs in an isolated git worktree/branch — safe from shared-checkout clobbering; you then compare attempts and merge the winner. |
Naming drift across surfaces. Cursor's docs describe the built-ins as
explore, bash, and browser, while Task-tool schemas expose
surface-dependent values (generalPurpose, explore, shell, browser-use,
best-of-n-runner, review specialists…). Read the live subagent_type enum
off the Task tool rather than assuming this table's names — the roles are
stable, the labels drift.
Custom subagents and the missing-type fallback. Custom subagents (project
.cursor/agents/, user ~/.cursor/agents/, or plugin-provided) show up as
their own subagent_type values — worth defining when a role repeats across
runs (a verifier, a docs researcher) so its prompt, pinned model, readonly,
and is_background defaults live in one file. Two verified gotchas
(staff-confirmed on the forum, not in docs): new agent files register only
after a Cursor restart, and a type missing from the enum is not permission
to skip the role — run it as generalPurpose with the role's instructions
inlined in the worker prompt (passing the intended model on the call)
instead. Nesting note: the platform itself now allows the main agent and its
direct subagents to launch subagents (one extra level, no deeper — the SDK
docs state the same cap); this skill still keeps fan-out orchestrator-only as
policy, because nested fan-out hides work from the manifest and the
verification gate.
Picking the model per slice (cost / speed routing)
Model choice is a cost/speed lever — route it, don't put every slice on a
frontier model:
- Scouting / decomposition / read-heavy exploration → the cheap, fast model.
Cursor's built-in
explore and search subagents already default to the
Composer fast family (e.g. composer-2.5-fast) for exactly this: fast, cheap,
and tuned for codebase understanding and tool use — so read waves are cheap by
default and you often need not set model at all. To pin it, pass
model: "composer-2.5" on the Task worker, or set the model field
(inherit | a model ID) on a custom .cursor/agents/ subagent. For
lightweight short-context slices on the GPT side, gpt-5.6-luna at a higher
effort is the cost-per-unit-of-work champion — but never give Luna
long-context reads (its recall collapses on 256K+ contexts per OpenAI's own
MRCR tables; route big-file slices to a stronger tier or chunk smaller).
This is the entropy-reduction workhorse.
- Per-model options ride in brackets on the model ID (documented syntax):
gpt-5.6-sol, claude-opus-4-8[effort=high,context=300k],
composer-2.5[fast=false], composer-2.5[] (empty brackets pin the
standard, non-fast variant). Use [effort=…] instead of guessing separate
"thinking" slugs.
- High-stakes verification, synthesis, or a multi-model panel → stronger
reasoning, chosen deliberately (e.g. a frontier tier at
effort=high,
escalating effort only for a slice that stays unresolved). For a
user-requested or high-stakes multi-model panel, ask which models to use;
don't guess slugs (see "Multi-model fan-out").
- Otherwise honor a model the user named; if a requested model is unavailable,
say so rather than silently substituting.
Caveats: availability varies (Max Mode, plan, or admin restrictions can force a
fallback to a compatible model; legacy request-based plans without Max Mode run
subagents on Composer regardless of model configuration); slugs drift, so
read them off Cursor's model picker rather than hardcoding volatile ones;
inherit can be unreliable in some surfaces (omit model to inherit). When a
custom agent's model matters, pin it in the frontmatter and pass the
matching model on the Task call — the field has been ignored under some
conditions (documented fallbacks plus confirmed bug reports), so if a worker's
output quality looks off, consider that the intended model may not have run.
Respect the user's cost and model preferences over any default here.
For review/audit slices, Cursor also exposes specialized subagents when available
(e.g. bugbot, security-review, ci-investigator, ci-watcher) — prefer them
for those slice types.
Multi-model fan-out (panel + synthesize)
The default fan-out runs each slice on one model. A high-stakes slice —
an architecture/design call, a risky correctness question, a security or audit
pass, or a key research synthesis — fan the same slice out to a panel of
different models, then act as judge and synthesizer.
Reconcile; don't concatenate. Label CONSENSUS (2+ models agree) vs
lone-model findings, resolve contradictions, dedupe overlap, and carry each
claim's confidence into one answer. The synthesis is where most of the value
lives — per OpenRouter's Fusion research, roughly three-quarters of the gain
comes from the synthesis step, not the model diversity — so invest there rather
than stapling outputs together. Evidence: a panel of independent models + a
judge + a synthesizer matched and surpassed a single top-tier model on hard,
deep-research problems (OpenRouter, Surpassing Frontier Performance with
Fusion, openrouter.ai/blog/announcements/fusion-beats-frontier/).
Run it in Cursor: pass a different model to each sibling Task worker on
the same slice, launch them in one message (parallel; backgrounded where
the surface supports it), then synthesize. This is the high-stakes end of model routing —
the orchestrator picks frontier models only when the user asks for a multi-model
pass or the slice is explicitly high-stakes — so ask which models to use;
don't guess slugs. (The cheap end — routing scouting and read-heavy waves to
Composer 2.5 — is under "Picking the model per slice.")
Caveat: a panel multiplies token cost (you pay every worker) and adds latency —
reserve it for high-stakes slices, not routine ones. The adversarial multi-model
review (a panel of reviewer models + one synthesized verdict) is this same
pattern applied to code review. For grading panels specifically, judges drawn
from disjoint model families beat a single frontier judge on human agreement
while cutting self-preference bias and cost (PoLL) — the family diversity, not
the head count, is what does the work.
Generate-and-filter & tournaments
For open-ended ideation or "produce the single best X", generate several
candidates and filter — don't trust one attempt:
- Cheap filter first. Gate candidates through a near-ground-truth check
(tests, schema/exec, dedup/clustering) before spending judge tokens —
generation is cheap, judging is not. (AlphaCode's shape: generate many → filter
~99% by tests → cluster → submit a few.)
- Selection ladder, not all-pairs. Dedup/cluster → shortlist → pairwise-judge
only among finalists. A naive O(N²) tournament spends most of its tokens
comparing also-rans.
- Use
best-of-n-runner for competing implementation attempts (each in its
own git worktree), then inspect/test/merge the winner yourself.
- For high-stakes judgment calls, the multi-model panel (above) is the
generate-and-filter of verification.
- Budget check. At equal cost, k independent attempts + a majority vote or
cheap filter usually beats critique/debate loops — benchmark any iterative
loop against that baseline before paying for it.
Worker prompt = the contract
A worker cannot ask follow-up questions. Under-specified prompts drift silently.
Write each as if you get one shot. Every worker prompt includes:
- Overall goal (context only — "don't try to own the whole goal").
- This worker's exact slice (the disjoint range / area / paths).
- Where to look (dirs, files, data ranges, which MCP/tools to use).
- What to return — the structured handoff from
references/handoff-format.md.
- Self-verify rules — cite-or-drop every claim, tag confidence
(
high|med|low), and state what you could not verify. Analysis workers:
"read COMPLETELY (N lines) and report the count." Research workers: "use live
sources, not training data."
- Scope boundaries — what's explicitly OUT of scope (so it doesn't overlap a
sibling worker).
- Return only the handoff, and keep it a digest. The worker's entire final
message becomes your context — tell it to return only the structured
handoff, capped at roughly 15 findings with one-line evidence each, and to
write any large artifact (tables, logs, full lists) to a file and cite the
path instead of pasting it inline.
- Edit workers: you are not alone. For implementation slices, warn the
worker that siblings may be active: own only the listed paths, don't revert
others' changes, and never spawn its own subagents (only the orchestrator
fans out).
See references/handoff-format.md for the copy-paste worker handoff template and
references/examples.md for a full worked run (the health-coach research job in
the screenshots) plus reusable decomposition recipes.
Parallel writes (the local gotcha)
Local subagents share one working directory. Concurrent writes to overlapping
files clobber each other.
- Parallel reads/research/analysis: always safe. This is the default use.
- Disjoint edits you keep all of: use regular workers only when path sets are
strictly disjoint, or do the edits serially after the research waves finish.
- Competing attempts at the same task: use
best-of-n-runner (each in its
own git worktree/branch), then inspect, test, and merge the chosen result
yourself — worktrees prevent clobbering, not the merge.
Waves instead of cloud orchestration
Waves is deliberately the replacement for cloud fan-out, not a stepping
stone to it. Local subagent runs cover the whole workload this skill targets —
research, audits, data analysis, exploration, and bounded implementation — with
isolation, parallelism, resume, and verification, and none of the cloud setup
(separate VMs, API keys, runtimes). The lessons worth keeping from the Cursor
team's cloud orchestrate plugin are already folded into this skill: planners
plan, workers hand off up, no cross-talk, disk/git as the durable medium. Do
not spawn cloud agents for a waves run; if a run truly outgrows one machine
(days-long fleets, PR-per-task pipelines), that is a different tool choice for
the user to make — say so and let them decide.
Cursor subagent mechanics here were checked on 2026-07-19 against current
docs, changelogs, and staff forum replies: parallel Task calls in one message
are the documented fan-out; is_background frontmatter is the documented
background switch while per-call run_in_background is live but undocumented
and absent on some surfaces; completed subagents persist resume checkpoints
(CLI release 2026-07-06); nesting is capped at one extra level; read-only
mode blocks all MCP; there is no documented concurrency cap, but staff report
~40 concurrent workers can overwhelm the extension host — batch into waves.
The details most likely to drift: Task parameter names, the subagent_type
enum, and model-ID bracket options. Re-verify those if they matter to your
run.
Checklist