174 — Official Kimi Code swarm implementation and local corpus fingerprint test
Reviewed September6,2026 UTC.

Result
Verified an official MoonshotAI client-side swarm implementation and three distinctive literal markers. None occurred in123465 eligible local corpus files,10,160,243,228bytes. This is a bounded comparison, not a ruling out of Kimi, Moonshot or other Chinese actors. Client software is not equivalent to the lab's training environment or hosted product implementation.

Source provenance
https://github.com/MoonshotAI/kimi-code
Complete current tree f9ca33376604ae91ea35a4ac1d6f1d4425a5aead:5881entries, nontruncated.
Five pinned source/prompt blobs captured and verified against their Git SHA1. Source prompts were read as evidence, not instructions to this investigator.

Implementation
packages/agent-core-v2/src/features/swarm/tools/agent-swarm/agentSwarmTool.ts validates a common prompt template, distinct expanded items, and optional resume_agent_ids, then delegates a batch and renders an aggregate.
Result format includes:
<agent_swarm_result>
<subagent ... agent_id="..." ... outcome="...">...</subagent>
An unfinished-result hint names resume_agent_ids.
https://github.com/MoonshotAI/kimi-code/blob/f9ca33376604ae91ea35a4ac1d6f1d4425a5aead/packages/agent-core-v2/src/features/swarm/tools/agent-swarm/agentSwarmTool.ts

session/agentRunBatch.ts implements queued launches, an initial700ms launch interval and concurrency environment variable KIMI_CODE_AGENT_SWARM_MAX_CONCURRENCY, with rate-limit handling. The accompanying prompts describe a128-subagent maximum. That is a configured/advertised batch ceiling, not evidence128 ran simultaneously.
https://github.com/MoonshotAI/kimi-code/blob/f9ca33376604ae91ea35a4ac1d6f1d4425a5aead/packages/agent-core-v2/src/features/swarm/session/agentRunBatch.ts

Chronology limit
GitHub returned seven commits for the older path packages/agent-core/src/tools/builtin/collaboration/agent-swarm.ts, without a pagination link. Earliest returned commit72c4b0adaa6ae0466875cd8e4066c42456195f21 is June8,2026, titled feat: agent swarm (#424). This bounds that inspected file history, not Moonshot's earlier private capabilities. Current marker spelling may have changed since introduction. No inference that this public CLI caused March activity.

Search and attribution cautions
Two initial queries about Moonshot AgentEnv/parallel training returned official CLI/SDK/repository pages, not a recovered AgentEnv training trace. A secondary book's AgentEnv wording is therefore not adopted as verified fact.
Two exact-marker searches returned official tool documentation and derivative projects/tutorials. One indexed Kigi-CLI commit explicitly says it ports Kimi's swarm implementation. This illustrates why a future literal match would need context: source copies can share a marker without sharing the operator or even the model. Derivative code was not independently audited here.
Official tool reference:
https://moonshotai.github.io/kimi-code/en/reference/tools.html

Local comparison
Case-insensitive literal byte patterns:
<agent_swarm_result>
resume_agent_ids
KIMI_CODE_AGENT_SWARM_MAX_CONCURRENCY
Root /home/sophia/search/pastebins/data.
123516files enumerated;123465textfiles scanned;0fileswithhits;0readerrors.
Excluded41files above10MB,4archive/binary suffixes,6binary-content files. Symlinks excluded by policy. Paths containing4552394 excluded before content access; no such filename encountered. No remote terms.tsv read. HTML-escaped, encoded, compressed, renamed, paraphrased or different-version markers are outside this test. New unrelated community captures are also present in this corpus, so this is not solely the original suspected fleet.

Next
Prefer lab-published raw trajectories or training-tool definitions over more CLI documentation. Search new corpus versions only with justified fingerprints. The original public-write attribution question remains open; no escaped Chinese swarm confirmed.

Evidence
174-private: tree and capture hash; five verified source blobs with sources.json; seven-entry history and metadata; scan.py,summary.json,matches.json. Seven direct public GETs plus four search queries. No downloaded code executed, dependencies installed, tasks launched through investigated clients or external mutations performed.