AI agent safety
For dashboard-ai.html · last checked 19 October 2026
AI Agent Disclaimer
AI agents generate responses based on the instructions, data, model, and tools available to them. They may produce inaccurate, incomplete, outdated, or misleading information.
Review AI-generated outputs before relying on them or sharing them with others. Verify important figures, calculations, interpretations, and recommendations against authoritative sources.
AI agents must not be treated as the sole decision-maker for consequential employment, legal, financial, medical, compliance, or other high-impact decisions.
The actions an agent can perform depend on the tools, permissions, and integrations configured in the application. In this version of the kit, agents have no tools, integrations or external actions: they return text. The one exception is Set up the view for me: it can suggest dashboard settings (filters, period, measure, chart). These change only when you click Apply, and you can undo them. See "What agents can and cannot do" below.
You are responsible for configuring and using agents appropriately and ensuring that their use complies with applicable laws, organisational policies, and data protection obligations.
A shorter version is shown in the dashboard under the agent buttons and in the agent editor.
What an agent actually is in this kit
An agent is a name plus written instructions. When you click Run agent, the dashboard sends one request to your chosen AI provider. That request contains:
- System instructions. These are fixed and the same for every agent. They tell the AI to use only the figures provided, not to invent numbers, to say plainly when something can't be determined, and to answer in plain English in under 220 words. They are instructions to the AI, not a technical control: a model can still ignore them or make mistakes.
- The agent's instructions (the "TASK").
- A summary of the current view, calculated by the dashboard. It contains only the columns you allow (see DATA-PRIVACY-AND-MEMORY.md).
The provider sends back text, which is shown on screen. That's all.
Agent settings that exist
| Setting | Exists? | Notes |
|---|---|---|
| Name | Yes | Up to 40 characters for your own agents (60 for imported) |
| Instructions | Yes | Up to 2,000 characters |
| Provider and model | One setting for all agents | Set in Set up connection |
| What the agent sees | One setting for all agents | Summary or Summary + 20 example rows, plus the column choices |
| Output format | Through instructions only | Ask for it in the instructions, e.g. "as a table" |
| Tools / actions | Not implemented | Agents can't take actions outside the dashboard |
| Change the dashboard view | Only through Set up the view for me | It suggests existing settings; nothing changes until you click Apply; Undo is available |
| Memory | Not implemented | Each run starts fresh |
| Permissions / approvals | Not needed in this version | There are no actions to approve |
What agents can and cannot do (verified in the code and by tests)
| Can an agent… | Answer |
|---|---|
| Send emails or messages | No. "Draft an update email" only writes text for you to copy |
| Change, delete or save your files or data | No |
| Change the dashboard's filters, period, measure or the "Build your own chart" settings | Only after you click Apply in Set up the view for me |
| Add new chart types, target lines, formulas, colours, layout or data | No. The built-in AI can only choose from settings that already exist. See "Bigger changes" below |
| Call other websites or systems | No. The safety lock blocks all connections except the provider addresses listed in it |
| Run code or workflows, publish content | No |
| See columns you've hidden, other files or other tabs | No. Only the summary in the preview is sent |
| Remember earlier runs or questions | No |
| See your API key | No. The key goes in a request header to the provider, not in the message |
Because agents can't act, the main risks are wrong answers and sending data you didn't mean to share. Both are covered below.
Guardrails to build into your own instructions
- Purpose: one clear job per agent, e.g. "Find the three weakest outlets this month".
- Data boundary: "Use only the breakdowns provided. Do not assume anything about data not shown."
- No invention: "If a number is not in the data, say it is not available."
- Facts vs assumptions: "Label suggestions as suggestions."
- Missing information: "Say what extra data would be needed to answer properly."
- People: "Talk about patterns, not individuals." Don't ask agents to judge named people.
- Escalation: "If the result could affect someone's job, pay or safety, say it needs human review."
Keep secrets and personal details out of instructions. The dashboard refuses to save or import instructions containing text that looks like an API key. Instructions are sent to the AI provider and are included in files you save.
High-impact decisions
Don't use agent output as the only basis for decisions about hiring, firing, pay, performance, credit, legal or medical matters, or compliance. Use it to point you to where to look, then check the data and involve the right people. In Singapore, if personal data is involved, the Personal Data Protection Act (PDPA) may apply (see DATA-PRIVACY-AND-MEMORY.md).
Importing agent packs safely
Before importing an agent pack: Review its instructions and data requirements. Do not import a pack from an untrusted source without inspecting its contents. Never include API keys, passwords, or other secrets in an agent pack. In this kit, packs contain written instructions only: they cannot add connections, code or permissions.
What a pack is: a .agents.json text file. You can open it in any text editor before importing.
What the importer reads (verified): only each agent's name and task text.
- Every other field is ignored, including any provider, address, key, script or code.
- Names and instructions are shown as plain text. They never run as code or web content.
Checks on import (tested): - Size: files larger than 300 KB are refused. - Format: invalid JSON, and packs with no agents, are refused with a message. - Number of agents: at most 30 agents are read per pack. Names are cut to 60 characters and instructions to 2,000. - Keys: agents whose name or instructions look like an API key are left out, and you're told how many. - Confirmation: before loading, you see the pack's name and its agents' names, and you confirm. - Replacing: if a pack with the same name is already loaded, you're warned that it will be replaced. Your own agents are never overwritten. - Reading instructions: to read an agent's full instructions, hover over its button, or select it and open the preview.
Remaining risk. A pack's instructions could try to manipulate the AI, for example "ignore the rules", to produce a misleading answer. They cannot make the dashboard send more data, connect anywhere new, or take actions. Read a pack's instructions before relying on its answers.
Bigger changes: the AI can't fully customise the dashboard
The built-in AI cannot redesign the dashboard or add features. Set up the view for me only picks from the filters, periods, measures and chart options that already exist. When you ask for something those settings can't do, such as a target line, a new chart type, a calculated column or new colours, it says so under "Can't be done with the built-in settings".
For bigger changes, the kit includes "PROMPTS - copy and paste into Claude.txt": - You give your own AI assistant (e.g. Claude or ChatGPT) the dashboard file and a prompt. - It writes a changed version of the file. - This happens outside the dashboard, with your own assistant. Results depend on the assistant, are not tested by FPA for every request, and may need several tries. Always keep a copy of the original file, and check the numbers afterwards.
Recommendations not yet implemented
These are possible future improvements, not current features: - Showing every agent's full instructions in the import confirmation (today: names only; instructions on hover or in the preview). - An option to let a pack's author sign it so you can check where it came from.
Also in this guide set: