Compare evidence from supplied sources or authorized project documentation. External web search and web fetch are currently unavailable in Veryfront Agent. The request must include the source text or identify the project files needed.
Agent-facing flow
- Load the Veryfront router skill:
veryfront__load_skill({ "skillId": "veryfront" }). - Load this page as the canonical docs resource for the task.
- If the request is broad, ask exactly one short
veryfront__form_inputwith one requiredbrieffield, then continue from the submitted brief. Ask for the source material when it is missing. - Gather evidence from supplied source material, authorized project files, project documentation, and available internal tools. Explain any evidence you cannot obtain. Do not bypass the external-search restriction through another tool or agent.
- Separate facts from inference and cite sources or file paths for claims that matter.
- Summarize the findings and the limits of the supplied evidence before recommending a next step.
- Save a report with
veryfront__create_fileorveryfront__update_fileonly when the user asks for a file or gives an exact path.
Guardrails
- Do not load a removed
researchsub-skill. - Do not treat the starter label, generic starter prompt, or storage location instruction as the research topic.
- Do not fall back to generic project inspection when the user asked for a specific research topic.
- Do not self-delegate on the first turn unless isolation materially improves the sourced research path.
- Do not create app, page, route, or chat UI files for a research-only request.
- When a report file is requested, write the complete findings and verify the saved content before reporting its path.
- Do not call a report grounded when the saved artifact lacks stable links, file paths, or tool references for material claims.
Example pattern
This example shows a report's structure and evidence questions. Replace them with findings from the supplied sources. Do not invent citations.
<!-- research/agent-tool-approval-patterns.md -->
# Research: Agent Tool Approval Patterns
Research brief: Compare practical patterns for deciding when an AI agent can call tools automatically and when it should ask for human approval.
## Questions
- Which tool actions are safe to run automatically?
- Which actions require user approval because they change external state?
- What evidence should the agent show before asking for approval?
- How should failures and partial tool results be reported?
## Evidence to gather
- Project agent instructions and existing tool-access rules.
- Product docs for skills, resources, tools, and agent creation.
- Additional source text supplied by the user when the project evidence is not enough.
## Resources
- `AGENTS.md`: Project-level rules for tool use and verification.
- `/agent/guides/create-agent`: Current project-agent creation and tool-selection guidance.
- `/cloud/mcp/tools/create-agent`: MCP schema for creating project agents.
## Output
- A short recommendation for default approval boundaries.
- A risk list for destructive or externally visible tool calls.
- Verification gaps where the project needs more product telemetry or policy detail.veryfront__create_file({
"project_reference": "<project>",
"path": "research/agent-tool-approval-patterns.md",
"content": "# Research: Agent Tool Approval Patterns\n\nResearch brief: Compare practical patterns for deciding when an AI agent can call tools automatically and when it should ask for human approval.\n\n## Questions\n\n- Which tool actions are safe to run automatically?\n- Which actions require user approval because they change external state?\n- What evidence should the agent show before asking for approval?\n\n## Resources\n\n- `AGENTS.md`: Project-level rules for tool use and verification.\n- `/agent/guides/create-agent`: Current project-agent creation and tool-selection guidance.\n- `/cloud/mcp/tools/create-agent`: MCP schema for creating project agents.\n\n## Output\n\n- Recommendation for default approval boundaries.\n- Risk list for destructive or externally visible tool calls.\n- Verification gaps.\n"
})The MCP call saves the Markdown in content at path. Include stable source references before final notes or open questions.
Output shape
Return the answer, supporting project evidence, recommendations, and verification gaps. Include the saved path when the user requested a report file.