Kimi Work Teardown: Goal Mode, WebBridge & 300 Parallel Agents

Moonshot AI's Kimi Work is a desktop AI agent system built for professional researchers — financial analysts, consultants, academics — not general users. While the company's ARR tripled from $100M to $300M in three months (Guangfa Securities tracking), Kimi Work is often dismissed in mainstream reviews as "just another desktop agent" because evaluators apply general-purpose criteria to a specialized research engine.

Product Origins: From Kimi Code to Research Engine

Kimi Work reuses the Kimi Code local-agent foundation (100k+ daily active developers). The Beta Mac/Windows clients were built in one week, producing 50,000+ lines of code with 92% AI-generated — demonstrating Moonshot's confidence in agent-built agents. The underlying model, Kimi K2.6/K3, supports 13-hour continuous coding, 300 parallel sub-agents, and 4,000+ tool calls per task. Architecture: model inference runs in the cloud; file I/O, browser control, and command execution run locally, keeping data and login state on-device.

Four Core Capabilities

Goal Mode (launched June 18, 2026) : User defines goal, acceptance criteria, constraints, and budget. Agent runs a persistent plan→act→verify→iterate loop 24/7, surviving sleep/meetings/commutes. Comparable to OpenAI Codex /goal but executed in a local desktop environment.

WebBridge : Local bridge service + browser extension lets the agent operate the user's already-logged-in Chrome/Edge — navigate, click, scroll, fill forms, scrape data. On the BrowseComp web-operation benchmark, Kimi K3 scored 91.2, ranking first among tested models.

Cron Scheduled Tasks : Built-in scheduler triggers agents for morning briefings, end-of-day summaries, overnight data processing. Includes a "Keep Computer Awake" toggle to prevent sleep-mode interruption. Free tier: 2 scheduled tasks; paid tiers add more.

300 Sub-Agent Parallelism (Agent Swarm) : A single instruction spawns up to 300 parallel sub-agents, each with independent context, aggregating results into a coherent deliverable (PPTX, Word, PDF, Excel, website) saved to local folders. Enables simultaneous financial-report reading, news scraping, chart generation — throughput impossible for single-agent mode.

Data Moat: Native Professional Sources

Kimi Work pre-integrates A-share, Hong Kong, and US market/securities data (Yahoo Finance, World Bank, Tonghuashun), corporate intelligence (Tianyancha), and academic databases (journals, preprints, theses, patents). Researchers can dispatch 300 agents to scan hundreds of papers, extract methodologies, map controversies, and produce structured literature reviews. Long-context strength: K2.6 supports 256K tokens; K3 extends to 1M tokens, enabling ingestion of entire financial-report stacks in one pass.

Critical Weaknesses

Speed : K3 outputs ~62 tokens/second, below peer median; latency compounds in long, multi-agent tasks.

Hallucination Rate Increase : Artificial Analysis's AA-Omniscience benchmark shows K3 hallucination rate rose from 39% (K2.6) to 51%. Definition: proportion of confidently wrong answers among (wrong + partially correct + not attempted). Accuracy improved 33%→46%, but when uncertain, K3 hallucinates more — risky for unsupervised contract drafting, CRM entry, client copy.

Long-Horizon Memory Loss : SoHu testing found attention dilution after 30 rounds, forgetting after 50+. Engineering test: 92% accuracy at round 10, plummeting to 63% at round 15 — context entropy overflow, not model failure. Moonshot's Memory Space caps at 50 entries × 500 chars; cross-session preference retention works, but intra-session coherence requires manual "summarize every 25 rounds" workarounds.

Data Leak Incident : April 2026 saw cross-account resume-data leakage; security/compliance gaps remain.

Compute Crisis: 48 Hours from Launch to Subscription Freeze

July 16, 2026: Kimi K3 launches — 2.8T parameters, 1M token context, #1 Frontend Code Arena, #4 Intelligence Index. ARR hits record single-day spike; enterprise API migration begins. July 19: Moonshot pauses all new C-end subscriptions; 48-hour request volume neared cluster capacity. All compute reserved for existing subscribers; rights unbundled (Web/App/Work vs. Kimi Code) for finer allocation. Three implications: (1) Model capability is no longer the sole moat — GPU supply is; (2) Growth bottleneck is supply, not demand — analysts (CIC 灼识) call this a structural valuation support; (3) Users depending on Kimi Work for production need fallback plans until capacity stabilizes.

The article concludes that Kimi Work's "overlooked" status stems from choosing a path hard to benchmark: fusing long-horizon execution, cluster collaboration, local data, and professional sources into a researcher-focused desktop app. Imperfect — speed, hallucination, memory, compute all need work — but precisely therefore deserves rigorous teardown, not a one-line dismissal.