Kimi Agent Swarm is Kimi’s parallel multi-agent mode for work that splits into independent branches, such as source collection or batch processing. Parallelism does not automatically make a task faster, cheaper or more accurate than standard Agent.
ACCESS OBSERVED / EXECUTION NOT RUN. The signed-in interface exposed the Agent Swarm/K3 Swarm entry on August 5, 2026. We submitted no task, verified no runnable membership or credit allowance, consumed no credits and observed no sub-agent work. This page contains no independent performance result.
The direct answer
Use Agent Swarm only when a task has many separable work units and you can define one shared source policy, output schema and acceptance test. Standard Agent is usually the clearer first choice for a short answer, a single document, a tightly sequential decision or a task where every step depends on the one before it.
Kimi says Swarm can coordinate up to 300 sub-agents and more than 4,000 tool calls, and reports about 4.5x faster execution than a single-agent workflow in suitable scenarios. These are Moonshot AI product claims, not KI AI Team measurements. Our planned same-prompt comparison has not run.
| Question | Current answer |
|---|---|
| Did the signed-in interface expose Agent Swarm/K3 Swarm? | Yes, observed on August 5, 2026 |
| Was access to a completed Swarm task verified? | No |
| Was a membership tier or runnable credit allowance confirmed? | No |
| Was a task submitted? | No |
| Were sub-agents or tool calls observed? | No |
| Were credits consumed? | No test consumption was initiated or measured |
| Is the reported 4.5x speed figure our result? | No; it is an official vendor-reported figure |
| Has standard Agent been compared with Swarm here? | No; both comparison conditions are NOT RUN |
What Kimi Agent Swarm is
Kimi’s Agent Swarm documentation describes an orchestrator that decomposes a complex objective and assigns parts to specialist sub-agents. Those agents can work in parallel, return intermediate findings and contribute to a combined deliverable. Kimi currently identifies the surface as powered by Kimi K3/K3 Swarm.
The useful distinction is the shape of the work:
- Standard Agent: one managed agent workflow handles the request and uses available tools as needed.
- Agent Swarm: an orchestrator can divide the request among many concurrent specialists and merge their work.
Parallelism creates capacity, not automatic correctness. If many workers use the same weak source, misunderstand the same field or produce overlapping rows, the final file may be faster but not better. A fair evaluation must therefore count missing sources, unsupported claims, duplicates and schema errors alongside wall time.
For the ordinary product surface, see the separate Kimi Agent guide. For the underlying model profile, see Kimi K3. This page is intentionally limited to Swarm access, task fit and a controlled comparison.
Official claims versus our observations
| Statement | Evidence type | Independent status |
|---|---|---|
| Up to 300 sub-agents | Current Kimi documentation | Vendor claim; not observed here |
| More than 4,000 tool calls | Current Kimi documentation | Vendor claim; not observed here |
| Approximately 4.5x faster in suitable comparisons | Current Kimi documentation | Vendor-reported comparison; not reproduced here |
| Powered by Kimi K3/K3 Swarm | Current Kimi documentation | Product statement; backend identity not independently inspected |
| Available to named paid membership tiers | Current Kimi documentation | Account-specific executable eligibility not verified |
| Swarm tasks consume significantly more credits than standard Agent | Current Kimi documentation | No before/after credit reading exists here |
| Agent Swarm/K3 Swarm entry visible in a signed-in web interface | Dated KI AI Team access observation | Observed August 5, 2026 |
| A Swarm task can complete on the observed account | No evidence | NOT VERIFIED |
| Swarm improved speed, source coverage or quality | No comparison outputs | TEST NOT RUN |
Seeing an entry point proves only that the interface exposed it at that time. It does not prove that the account could start or finish a task, that a specific number of workers would be allocated, or that the mode would consume a predictable amount of credit.
Standard Agent or Agent Swarm?
| Task characteristic | Standard Agent is the safer default | Agent Swarm may fit |
|---|---|---|
| Number of independent work units | One or a few | Dozens or hundreds |
| Dependency pattern | Steps must happen in sequence | Branches can proceed independently |
| Source universe | Small or exploratory | Large, explicit and allowlisted |
| Output | One concise answer or document | A batch, wide evidence map or multi-part deliverable |
| Quality control | One reasoning chain is easy to audit | Row-level validation can detect omissions and duplicates |
| Cost sensitivity | Minimize credit use | Extra credits are authorized and justified |
| Coordination overhead | Would dominate the task | Is small compared with parallel work saved |
| Failure risk | One missing step can be corrected directly | Merge conflicts and inconsistent worker outputs can be checked systematically |
This is a decision framework, not a measured Kimi performance table. The Swarm column reflects the product’s documented design and general multi-agent trade-offs; it is not evidence that any particular Kimi task will improve.
Tasks that are shaped well for Swarm
Kimi lists large-scale retrieval, batch downloads, wide reading, long outputs, complex programming and office automation among suitable areas. In practice, the strongest candidates share four properties: the units are separable, the completion rule is explicit, each unit can be checked independently, and the final merge has a defined schema.
Examples include:
- extracting the same fields from a large, fixed set of official pages;
- checking a collection of files against one compliance checklist;
- implementing independent modules with a shared interface and a final test suite;
- preparing distinct sections of a report from non-overlapping source packets; and
- converting a folder of documents into a row-by-row inventory with validation rules.
These examples describe task shape. They are not claims that Kimi Swarm completed them in our account.
Tasks that are poor Swarm candidates
Avoid paying coordination overhead when the work is small or fundamentally sequential. Poor candidates include:
- rewriting one paragraph;
- answering a question from one short source;
- making a single irreversible account change;
- a calculation where every step depends on the previous result;
- a task with an undefined output format or no acceptance rule;
- a source search with no boundary, where many workers could amplify low-quality material; and
- work containing secrets or personal data that should not be distributed across unnecessary tool calls.
If the task cannot be divided on paper without ambiguity, adding more agents is unlikely to solve the specification problem.
How to access Kimi Agent Swarm
The current official route is kimi.com/agent-swarm. Kimi’s documentation also says mobile users can select K3 Swarm where the feature is available. Product labels and rollout can change, so these are documented steps rather than a promise of access for every account.
- Sign in to the official Kimi product.
- Open the Agent Swarm route or select K3 Swarm in a supported client.
- Check that the selected mode is visibly Agent Swarm/K3 Swarm before composing the task.
- Review the account’s current membership, available credits and any task estimate.
- Define the allowed sources, output schema, stop conditions and privacy boundaries.
- Submit only after the expected credit exposure is acceptable.
- Preserve the task tree, downloadable result, timing boundary and visible credit change if you intend to evaluate the run.
We completed only the interface-observation portion of this sequence. No task was submitted, so the later steps are not personal-use claims.
Membership and credits
Kimi currently documents Agent Swarm for the paid Moderato, Allegretto, Allegro and Vivace membership tiers. The company also warns that Swarm work consumes significantly more credits than standard Agent. Exact allowances and plan details are volatile; check the official membership pricing, membership credits page and Agent quota and billing guide while signed in.
We do not infer eligibility from the presence of a menu item, repeat a tier allowance as a permanent entitlement, or translate credits into tokens without documentation. The observed account’s plan and balance are intentionally not claimed here because they were not verified for execution.
The preregistered Standard Agent versus Swarm comparison
The research question is narrow: on one allowlisted official-source synthesis task, does K3 Swarm improve source coverage or wall time over standard Kimi Agent without disproportionately increasing unsupported claims, duplicate records or visible credit use?
It is not a general benchmark. One well-controlled task can reveal coordination behavior, but it cannot establish that Swarm is universally faster or better.
The source ledger gate
The full run requires a ledger of at least 40 manually verified official URLs. Each row needs a stable source ID, expected page title, product area, one exact fact oracle, destination URL and a limitation note. The ledger is not complete and is not published today. Therefore, even if runnable credit access appeared, the benchmark would remain ineligible to start.
A smaller or unverified source set may be used only as a separately labelled pilot. It cannot be silently promoted to the full comparison.
The fixed task
Both conditions receive the same request:
Using only the allowlisted official-source ledger supplied with this task,
produce one CSV row per eligible source. Use these columns in this order:
source_id, page_title, product_area, requested_exact_fact,
supporting_url, limitation.
Do not use a source outside the allowlist. Do not infer a missing fact.
If a page cannot be accessed, retain its source_id and write UNAVAILABLE in
the requested_exact_fact field. Return the completed CSV and a concise run
summary listing inaccessible sources and unresolved limitations.The prompt, ledger bytes and expected schema are identical for standard Agent and K3 Swarm.
Fair-comparison controls
- Use the same signed-in account, day, source universe, prompt and required CSV schema.
- Open a fresh conversation for each condition.
- Randomize the order with a recorded coin flip before either output is visible.
- Confirm the displayed mode at submission and completion.
- Do not send a correction or follow-up before scoring.
- Accept no outside source for the core score.
- Record start and end using the same timing boundaries.
- Capture a visible task tree and sub-agent count only if the product exposes them; never estimate hidden workers.
- Record credits before and after when the interface exposes comparable readings; do not infer token use from credits.
- Preserve failed, blocked and partial attempts instead of rerunning them invisibly.
If one condition sees different sources, instructions or account state, the pair is not comparable. The deviation is reported and the speed ratio is withheld.
Scoring rules
| Metric | Calculation | Why it matters |
|---|---|---|
| Source coverage | Eligible source IDs represented / eligible source IDs | Detects omitted work |
| Fact accuracy | Rows matching the preregistered oracle / scored rows | Tests exact extraction, not writing style |
| Citation validity | Supporting URLs resolving to their allowlisted destinations / cited rows | Detects broken or substituted citations |
| Unsupported claims | Count of factual claims not supported by the cited allowlisted source | Detects invention or source drift |
| Duplicate rate | Duplicate IDs or semantically duplicate rows / returned rows | Measures wasted parallel work and merge quality |
| Schema compliance | Rows with the six required columns and valid types / returned rows | Tests deliverable usability |
| Wall time | Submission timestamp to final downloadable deliverable | Compares elapsed time with identical boundaries |
| Visible credits | Comparable after-reading minus before-reading, if exposed | Records account-visible consumption without inventing token equivalents |
Coverage and accuracy are reported together. A mode must not gain an apparent accuracy advantage by returning only easy rows. UNAVAILABLE is correct only when the page was genuinely inaccessible and the source ID is retained; using it for a difficult but accessible fact counts as missing coverage.
No “4.5x” comparison will be calculated unless both runs finish under the same timing definition. If the standard run takes 20 minutes and Swarm takes 10, the observed ratio would be 2.0x for this task, not evidence about other workloads. If either run fails, we report the failure and raw elapsed time without manufacturing a multiplier.
Current comparison ledger
| Evidence item | Standard Agent | Agent Swarm |
|---|---|---|
| Mode entry observed | Not recorded for this comparison | Yes, entry observed August 5, 2026 |
| Executable plan/credit allowance verified | No | No |
| Fixed 40-source ledger ready | No | No |
| Task submitted | NOT RUN | NOT RUN |
| Final CSV | NOT AVAILABLE | NOT AVAILABLE |
| Source coverage | NOT CALCULATED | NOT CALCULATED |
| Fact accuracy | NOT CALCULATED | NOT CALCULATED |
| Citation validity | NOT CALCULATED | NOT CALCULATED |
| Unsupported claims | NOT COUNTED | NOT COUNTED |
| Duplicate rate | NOT CALCULATED | NOT CALCULATED |
| Schema compliance | NOT CALCULATED | NOT CALCULATED |
| Wall time | NOT AVAILABLE | NOT AVAILABLE |
| Visible credit change | NOT MEASURED | NOT MEASURED |
The absence of output is not a tie and not a zero. There is simply no comparison result.
Evidence required for a valid run
A publishable comparison must retain:
- the final manually verified source ledger and its SHA-256 hash;
- the exact fixed prompt and output schema;
- dated entry screens showing the chosen mode;
- start and completion timestamps using one clock and one boundary;
- task tree, worker count and tool-call information only when visibly exposed;
- both original downloadable CSV files without silent editing;
- a link-resolution audit and row-by-row oracle score;
- before/after credit readings when comparable values are visible;
- screenshots with account identifiers and unrelated data redacted; and
- a complete deviation and failure log.
The public benchmark pack will be linked only after both conditions are complete enough to interpret, the source licenses and privacy boundaries are checked, and secrets are removed. There is no fabricated download link on this draft.
Safety and privacy checklist
Parallel execution can multiply both productive work and mistakes. Before submission:
- use synthetic, public or explicitly authorized files only;
- remove API keys, passwords, private links and unnecessary personal data;
- allowlist domains when source provenance matters;
- prohibit publishing, messaging, purchasing or changing accounts unless separately authorized;
- require confirmation before any irreversible action;
- state a credit ceiling and stop rule;
- define how duplicates, inaccessible pages and contradictions must be represented; and
- review the final merged artifact, not only the orchestrator’s summary.
The planned comparison is read-only. It authorizes retrieval from a fixed official source list and creation of a local CSV; it does not authorize external publication or account changes.
Limits of this guide
- The observed interface entry is one dated access observation, not proof of completed-task access.
- We did not verify the account’s paid tier or executable credit allowance.
- No task tree, sub-agent count, tool-call count, deliverable, latency or credit change was observed.
- The 300-agent, 4,000-tool-call and 4.5x figures are vendor statements, not independent results.
- Official membership tiers, credits and rollout conditions can change.
- The planned benchmark covers one structured official-source task, not coding, office automation or every Swarm use case.
- A 40-source ledger is still small relative to the product’s maximum-scale claims.
- Wall time can vary with service demand, source availability, region and account priority.
- Credit readings do not reveal token use unless the product explicitly maps them.
- Better coverage could coexist with more unsupported claims or duplicates; no single metric is sufficient.
See the Kimi AI Test Lab for completed and not-run records, the testing methodology for our evidence rules, and Kimi Deep Research for the separate research workflow intent.
Frequently asked questions
Is Kimi Agent Swarm the same as standard Kimi Agent?
No. Kimi describes Swarm as a parallel architecture in which an orchestrator delegates work to specialist sub-agents. Standard Agent remains the simpler surface for many tasks. Our page does not claim that one is universally better.
Is Agent Swarm available on the free plan?
The official documentation checked on August 5, 2026 lists paid Moderato, Allegretto, Allegro and Vivace tiers. Account visibility and rollout can change, and seeing the entry did not verify that our observed account could execute a task. Check the current signed-in membership page.
How many agents does Kimi Agent Swarm use?
Kimi says the architecture can coordinate up to 300 sub-agents. That is a maximum product claim, not a promise that every task receives 300 workers. We observed no task tree and will not estimate a hidden count.
Is Kimi Agent Swarm really 4.5x faster?
Kimi reports approximately 4.5x faster execution in suitable comparisons. We have not reproduced that result. Our protocol will publish a task-specific ratio only if both controlled runs finish with identical timing boundaries.
Does Agent Swarm use more credits?
Kimi warns that Swarm tasks use significantly more credits than standard Agent. We did not submit a task or record a credit delta, so we cannot quantify the difference. Review current terms on the membership guide before execution.
What is the best first Swarm task?
A bounded, read-only batch is safer than an open-ended or irreversible job: use a fixed source list, one row schema, explicit missing-data labels and a small approved credit ceiling. That makes both quality and consumption easier to audit.
Has the Standard Agent versus Swarm comparison run?
No. The 40-source ledger is not complete, executable membership/credit access was not verified and neither task was submitted. Every performance field remains NOT CALCULATED.
Official sources and update policy
Product facts were checked on August 5, 2026 against Kimi’s Agent Swarm documentation, Agent overview, Agent features and limits, quota and billing guide, membership pricing and membership credits. The current Agent Swarm route is kimi.com/agent-swarm. Background model research is available in the Kimi K2.5 paper, but paper claims are not treated as results of this product comparison.
We will update the access record only from a dated signed-in observation and will add performance findings only after the fixed ledger, both condition outputs and row-level scoring evidence exist. Corrections follow our Sources and Corrections policy.