Long Horizon, Deep Research, Quality control & Agent Swarm
Most tasks should be Standard tasks. They're the fastest and the easiest to steer.
For bigger jobs, choose one of four workflows from the Task Parameters menu before you send the task:
| Workflow | Use it when | What it adds |
|---|---|---|
| Long Horizon | The job is large and easy to half-finish | A written plan, and a reviewer that must approve the result |
| Deep Research | You need a well-sourced answer | The same plan and review, with research guidance |
| Quality control (experimental) | How the finished files look matters | A review pass that opens the files and checks layout and writing |
| Agent Swarm | The work can be split up | A leader that hands parts to copies of itself, then has the answer checked |
Long Horizon
Use Long Horizon when you want the agent to stay disciplined over a larger piece of work.
What happens
- A clarify stage runs first instead of starting execution immediately.
- The agent asks clarifying questions if needed.
- Once ready, it writes a durable
PLAN.md. - The main agent executes against that plan.
- When it thinks it is done, it submits its work to a reviewer.
- The task only finishes once the reviewer approves.
Why use it
Long Horizon is a good fit when:
- the task is broad and easy to half-finish,
- quality matters more than speed,
- you want a visible plan before deep execution starts,
- you want review pressure before the final answer is sent.
What you see in Task Detail
The workflow panel shows:
- the saved plan,
- the latest submission,
- the current round,
- the reviewer verdicts for that round.
If a reviewer wants changes, the task keeps going instead of silently stopping early.
Deep Research
Use Deep Research when you need a well-supported answer rather than a quick overview. It uses the same clarify, plan, execute, and review structure as Long Horizon, with research-specific guidance.
What happens
- The task clarifies the goal, audience, scope, freshness, or deliverable when that is genuinely needed.
- It plans whether the task calls for deeper investigation of a specific subject, wider comparison across sources and viewpoints, or both.
- The agent researches authoritative primary sources when they are required, while also using credible independent reporting, forums, or user accounts when they provide relevant real-world context.
- It keeps you updated at meaningful research milestones instead of leaving you to wait without context.
- The final report follows your requested format. Otherwise, it starts with the conclusion, explains the important evidence and uncertainty, and ends with the sources actually used.
Why use it
Deep Research is useful when:
- you need an answer grounded in current documentation, reports, or public evidence,
- official information may be incomplete and independent reporting or user experience matters too,
- you need to compare competing claims, policies, products, or programs,
- you want a concise answer without losing the supporting evidence.
What you see in Task Detail
The workflow panel is labeled Deep Research and shows the same plan, submission, review round, and reviewer verdicts as Long Horizon. A rejected review sends the task back to strengthen the research or answer before delivery.
Quality control (Experimental)
Use Quality control when the way the result is presented matters as much as completing the underlying work. It uses the same plan-and-review structure as Long Horizon, but the reviewer concentrates on the experience of reading, viewing, or using the finished result.
What happens
- The task clarifies the request and writes a durable plan.
- The main agent completes the work and prepares the full reply it intends to send.
- It identifies every user-facing file or visual artifact it created or changed.
- The reviewer checks the reply from the user's point of view and opens the actual output files.
- Presentations, documents, PDFs, spreadsheets, webpages, canvases, and images are rendered or viewed when visual inspection applies.
- The reviewer checks the result against Meowbert's writing style guide and requests another revision if wording, flow, formatting, or design problems remain.
Why use it
Quality control is useful when:
- the final answer needs to be especially clear, natural, or easy to follow,
- a paper, report, or presentation needs a careful editorial pass,
- generated files must look polished rather than crowded or template-like,
- visual consistency, typography, spacing, and layout are part of the deliverable.
What you see in Task Detail
The workflow panel uses the Quality control label and shows the same plan, submission, review round, and reviewer verdicts as Long Horizon. A rejected review keeps the task working until a corrected version is approved.
Agent Swarm
Use Agent Swarm when a task is big enough that one agent working alone would be slow, lose detail, or settle on its first idea.
An Agent Swarm is one agent, the leader, that can make copies of itself. You talk to the leader exactly as you would to a standard task. Most turns stay ordinary turns: a request like "rewrite this email to sound less defensive" is simply answered. When the work warrants it, the leader makes one of three moves.
What the leader can do
- Split. The leader hands different parts of the work to lanes, copies of itself that start with everything it knows. By default the leader pauses until its lanes finish, or it can keep working on its own part in parallel while finished lane results arrive between steps. Lanes share short notes directed to their siblings, their parent, or both—and can wake a waiting parent or send an urgent interrupt when a finding or blocker needs immediate action. For example, a lane researching EU grants that finds a rule ruling out your company shares it so the other lanes and the leader take it into account.
- Diverge. For a judgment call, the leader asks several lanes the same question. They answer without seeing each other, so they don't just agree with the first idea. The leader then compares their reasoning: where they agree, where they disagree, and ideas only one lane had. It builds its answer from the strongest one.
- Check. Before a reply that used lanes is delivered, a fresh reviewer that has seen none of the leader's reasoning reads what you asked and the draft, and looks for what is wrong. The leader fixes what it finds and then replies.
Lanes can split their own part again when it turns out to be bigger than expected.
Some examples:
| Request | What the swarm does |
|---|---|
| "Find every EU and Portuguese funding program a 10-person climate-tech startup could apply to in 2027" | The leader looks around, then splits into one lane per family of programs. Lanes share eligibility rules they discover. A reviewer spot-checks deadlines before the table is delivered. |
| "Propose pricing for our B2B analytics tool" | The leader diverges into three independent proposals, compares their key decisions, settles the one real disagreement with a focused lane, and builds from the best proposal. |
| "Turn these research notes into a 40-page market report" | The leader writes a short outline of terms and shared numbers first, then splits the sections across lanes. The reviewer looks for seams between sections. |
Small models make weaker leaders. Use a swarm whose leader runs on a strong model.
Check review effort
A Check lane reviews the draft and its files with fresh context. The leader can choose between two review levels:
| Level | What it checks |
|---|---|
| Standard | Correctness, missing requirements, contradictions, factual accuracy, readability, and writing rules. This is also the level used for automatic checks. |
| Thorough | The same review criteria as Quality control, including rendered files, visual quality, audience suitability, and matters of taste. |
Ask the leader for a thorough check when presentation matters, for example: "Before sending the report, run a thorough check of the writing and rendered pages." The agent chooses the level when it starts the Check lane; the lane's side panel shows Standard review or Thorough review. Review effort is separate from model reasoning effort and the swarm's delegation Effort setting.
Review findings have priorities, ordered from highest to lowest:
| Priority | Meaning |
|---|---|
| P0 — Critical | An exceptional issue requiring immediate attention because it risks severe consequences. |
| P1 — High | A major correctness, usability, or requirement failure that needs correction before delivery. |
| P2 — Normal | A concrete issue that materially reduces accuracy, completeness, clarity, or quality. |
| P3 — Low | A smaller, actionable improvement with limited impact. |
Both Check levels report findings for the leader to address. A Thorough Check uses Quality control's review criteria, while the separate Quality control workflow retains its approval and revision process. Older Check results may still show must_fix or should_fix labels.
Talking to the swarm while it works
The leader stays available while lanes work. Send a message at any time: the leader can tell you how far each lane has got, steer a lane, stop one, or start more.
After a reply, lanes stay dormant for a few turns with everything they found. When you ask a follow-up such as "go deeper on the EIC programs", the leader picks up the lane that already did that part instead of starting over.
If a lane fails, for example because its model hits a rate limit, ask the leader to resume it after the problem is resolved. Failed lanes keep their context and can be resumed within the same dormant period. Stopped lanes and Check reviewers cannot be resumed.
Why use it
Agent Swarm is useful when:
- the work has parts that can be done at the same time, such as research across many sources or sections of a long document,
- you want to see where an answer is uncertain, not just one confident path,
- you want a reply checked by a reviewer that does not share the author's blind spots.
For tightly connected, step-by-step work, a standard task is usually better.
Configuring the swarm
Pick an Agent Swarm in the model list, or choose Agent Swarm in the Task Parameters menu. The swarm's settings start from its preset, and you can change them for this task:
- Lane model router, when your server has model routers: the leader can run some lanes on other models, for example a reviewer on a different model. Without one, every lane uses the leader's model.
- Effort is the quick delegation control. Balanced keeps the normal behavior; lower values make the leader wait longer before splitting, while higher values encourage earlier and wider delegation. The setting scales the spawn-related Advanced limits without changing budgets or the provider concurrency cap.
- Budgets are off by default. Turn them on to give the leader and all lanes one shared token budget and an optional time budget. A part is held back for putting the answer together and checking it. Each lane gets a share; near the end of its share it is told to wrap up, and at the end it reports what it has. Every lane always reports something.
- Advanced... holds the limits: how many lanes a parent can start, how deep lanes can split, how many run at once, steps per lane, automatic Check, how many turns lanes stay dormant, and a few more.
Tool access
Whatever tools you enable in the composer apply to the swarm too. If you enable Web Search, Memory, Computer Use, Subtasks, or Skills, the leader and its lanes receive those same selected capabilities.
What you see in Task Detail
Under the leader message that started lanes, a compact block shows one row per lane: what it owns, its status, its latest note and its steps. Sub-lanes are listed under their lane. Open a lane to inspect its conversation and live context usage in the side panel, switch to Shared notes for a chronological timeline of shared notes (for that lane or the whole swarm), open Details for its model, token spend, timing, and direction, or stop it to make it report what it has now.
The header shows Waiting for lanes while the leader is paused on active lanes, along with how many lanes are working, budget use when budgets are on, and the dormant lanes with a one-line summary each.
In the conversation, once the leader delivers its final answer, the work leading up to it, lane blocks included, folds into one Worked for… row above the answer. Files and any error notices stay visible.
Choosing the right mode
Use Standard task when:
- the task is straightforward,
- speed matters,
- one agent is enough.
Use Long Horizon when:
- you want a plan first,
- you want formal review before completion,
- you are worried about the agent stopping too early.
Use Deep Research when:
- source quality and current evidence matter,
- you need wide comparison, deep investigation, or both,
- you want the conclusion and the sources behind it in one organized result.
Use Quality control when:
- the writing and organization need an editorial review,
- the task produces visual or formatted deliverables,
- you want the output judged from the reader's or viewer's perspective.
Use Agent Swarm when:
- the work has independent parts that can run in parallel,
- you want independent answers compared rather than one path,
- you want a fresh reviewer to check the reply before it is delivered.
Important limits
- Workflow type is chosen at task creation time.
- After the task is created, the workflow type is fixed.
- You can still adjust normal task parameters like max steps or waiting behavior where supported.
- Enable Web Search in the composer when the research needs current web information. Other selected tools and sources remain available to the workflow too.
- Quality control is experimental and may take longer because revisions continue until the presentation-focused reviewer approves the result.