Kimi K2.5 Vision Coding and Agent Swarm Task Prompt

One-sentence takeaway

Giving Kimi K2.5 explicit goals, parallelizable subtasks, an evidence format, and completion checks makes it suitable for visual-to-code work, web reconstruction, research aggregation, and multi-agent workflows. For complex tasks, Agent Swarm can let the model decompose the work itself, but concurrency and result verification still need to be constrained.

Use cases

• Suitable tasks: image/video-to-code, visual debugging, web reconstruction, cross-source research, batch classification, and office deliverables.

• Unsuitable tasks: unbounded parallelism without verifiable criteria; high-risk write operations or sensitive data should not be handed to automatically generated sub-agents.

• Applicable model versions: Kimi K2.5 Thinking/Instant/Agent; Agent Swarm is beta/research preview, subject to the current product status.

• Applicable clients, agents, or APIs: Kimi.com, Kimi App, Kimi API, Kimi Code, and third-party OpenAI/Anthropic-compatible endpoints.

• Recommended reasoning levels and parameters: Thinking uses the official recommendation of temperature=1.0 and top_p=0.95; Instant uses temperature=0.6 and top_p=0.95. Prefer Thinking for complex visual/search tasks.

Ready-to-use content

You are a vision and engineering agent.

Goal: Based on the provided image/video/requirements, deliver a runnable and verifiable result.

Work in the following order:
1. First inspect the input and list confirmed visual facts, missing information, and acceptance criteria.
2. Break the task into independent subtasks that can run in parallel:
   - Visual structure/layout analysis
   - Asset and text extraction
   - Code implementation
   - Runtime and screenshot verification
   - Interaction/accessibility checks
3. Subtasks should return structured results only: {task, evidence, result, risks}.
4. Resolve conflicts during aggregation; mark unsupported content as unknown and do not guess.
5. Run the code, take screenshots, or run tests for verification; if the acceptance criteria are not met, continue fixing until the budget limit is reached.

Final output:
- Completed files/links
- Key visual and engineering decisions
- Verification commands and results
- Unresolved items and next steps

When using Agent Swarm, independent tasks such as “research 100 specialized fields” can be handed to automatic orchestration. Do not force the prompt to blindly generate a fixed number of sub-agents; first let the model determine whether the work can be parallelized.

Test/workflow steps

1. Prepare an image/video and target page, and record the input dimensions, text, interactions, and responsive acceptance criteria.

2. Test a single agent and Agent Swarm separately, keeping the tool set, context, and maximum number of steps fixed.

3. Save each subtask's prompt, result, sources, code diff, screenshots, and test logs.

4. Check whether subtasks are duplicated, whether key facts are omitted, and whether conflicts trigger incorrect voting; the main agent/rule checker performs final acceptance.

Original evidence and data

• Kimi officially says that K2.5 can use visual inputs for code, visual debugging, and web reconstruction; the model repository describes it as a native multimodal agent model.

• The official Agent Swarm design can dynamically create up to 100 sub-agents and coordinate up to 1,500 tool calls; the company says it can reduce execution time by up to 4.5× compared with a single agent.

• The official blog gives a visual maze case: the model used code execution/BFS, reported a path length of 113,557 steps, and provided visualization verification; this is a public example, not an independent test.

• K2.5 supports image/video inputs, Thinking/Instant modes, and Kimi Code; Agent Swarm is labeled beta/research preview in the official blog.

Boundaries of use

• The 100/1,500/4.5× figures for Agent Swarm are official capability/internal-evaluation descriptions; actual results are affected by tool limits, concurrency, failed retries, and task parallelizability.

• Model-generated decomposition does not mean every subtask is correct; evidence, a schema, idempotent writes, and final regression tests are required.

• Coordinates, text, and layout in visual outputs must be checked with image/DOM/code tools rather than relying only on textual descriptions.

• This template is a reusable prompt adapted from official task methods, not the complete official system prompt.

Source excerpt or observation (compliance-short quote only)

The official description of Agent Swarm is “scaling out, not just up” (a compliance-short quote).