Persona Agent¶
Thinklio Built-in Agent Specification Version 0.1 | March 2026
1. Purpose and Problem Statement¶
The Persona Agent is a document-grounded, role-specific assistant. It answers questions, guides decisions, and provides expert-level responses within a defined domain — drawing on a curated library of documents, policies, and reference material as its authoritative knowledge source.
It is the base pattern behind several named agents in the Thinklio catalogue. The distinction between a Coach Agent, an HR Agent, and an Onboarding Agent is not architectural — it is configuration. All three are Persona Agents with different libraries, different system prompts, different audience assumptions, and different governance profiles.
The problem the Persona Agent solves is the gap between institutional knowledge and the people who need it. An organisation's HR policies, coaching methodology, onboarding procedures, or compliance guidelines exist in documents that are hard to navigate, inconsistently applied, and invisible to the people who need them most. The Persona Agent makes that knowledge conversational, consistent, and always available.
Critically, the Persona Agent is bounded. It answers from its library. When asked something outside its domain, it says so clearly. It does not hallucinate policy. It does not speculate beyond its sources. This boundedness is what makes it trustworthy in sensitive domains.
2. The Library Model¶
The Persona Agent's knowledge lives in a library — a curated collection of documents that have been processed, chunked, and indexed for semantic retrieval.
A library is: - Scoped to an organisation (private) or to the platform (shared across all deployments) - Seeded by an admin before the agent is useful - Updatable: adding or replacing documents updates the agent's knowledge without reconfiguration - The primary determinant of the agent's usefulness — an empty library produces a shallow, generic agent
Documents in the library are not read verbatim. They are chunked into semantically coherent segments, each indexed with an embedding vector. When a user asks a question, the agent retrieves the most relevant chunks and reasons from them. The user does not see the retrieval process — they see a coherent, sourced response.
Source attribution is a first-class feature of the Persona Agent. Every substantive claim in a response is traceable to a specific document and section. Users can ask "where does that come from?" and the agent will cite the source. This is what separates a Persona Agent from a general chat assistant — its answers are verifiable.
3. UI Structure¶
Chat Tab¶
The primary interface. Natural language questions and answers within the agent's domain: - "What is our parental leave policy for casual employees?" - "How should I handle a performance improvement conversation?" - "What do I need to complete in my first week?" - "Can you walk me through the incident reporting process?"
Responses include inline source citations. Users can expand a citation to see the source document and section. A "I don't know" response is produced when the query falls outside the library's coverage, with a suggestion to contact the relevant team.
Library Tab¶
A view of the documents that make up this agent's knowledge base. Visible to users (not just admins) so they can see what the agent knows and browse source material directly. Each document shows: title, type, date added, and a relevance indicator (how often it contributes to responses).
Admins see additional controls: add document, replace document, remove document, reprocess.
Topics Tab¶
An auto-generated map of the topics the agent can answer questions about, derived from the library contents. Useful for new users who don't know what to ask. Organised by theme (e.g. for HR: Leave Entitlements, Performance Management, Workplace Safety, etc.).
4. Configuration¶
4.1 Admin Configuration¶
| Setting | Description |
|---|---|
| Agent name and persona | Display name and the persona the agent presents (e.g. "HR Advisor", "Your Onboarding Guide") |
| Domain description | One-paragraph description of what this agent covers, used in out-of-scope responses |
| Library assignment | Which library (or libraries) this agent draws from |
| Source attribution mode | always (every response cites sources), on_request (only when asked), never (for internal use cases where source visibility is undesired) |
| Out-of-scope behaviour | What the agent says when asked something outside its domain: decline_and_redirect (default), decline_only, or attempt_with_caveat |
| Redirect target | Who or what to redirect out-of-scope questions to (e.g. "contact HR at hr@org.com") |
| Confidence threshold | Minimum retrieval confidence below which the agent declines rather than guesses |
| Response formality | Formal / professional / conversational — affects tone without affecting content |
| Audience | internal_staff, management, new_employees, customers, mixed — adjusts vocabulary and assumed context |
4.2 User Configuration¶
Users cannot modify library content or agent behaviour. Their configuration is limited to: - Preferred response length (concise / detailed) - Whether to show source citations inline or collapsed
5. Agent Capabilities¶
| Capability | Description |
|---|---|
| Answer from library | Retrieve relevant chunks and synthesise a response with source attribution |
| Explain / expand | Given a previous response, provide more detail or a worked example |
| Compare | Given two policy options or scenarios, compare them from the library's perspective |
| Summarise document | Produce a plain-language summary of a specific document in the library |
| Check applicability | Given a scenario ("I'm a part-time employee who has been here 6 months..."), determine which policies apply |
| Escalate | Identify when a question requires human judgement beyond what the library covers and produce a structured escalation note |
| Browse topics | Return a list of topics the agent can help with |
| What's new | Surface recently added or updated documents in the library |
6. Governance Profile¶
The Persona Agent's governance profile varies significantly by instance. The three canonical instances have different risk levels:
| Instance | Domain | Risk level | Key governance concerns |
|---|---|---|---|
| Coach Agent | Wellbeing, performance, personal development | Medium | Duty of care — must not substitute for professional advice |
| HR Agent | Policies, entitlements, procedures | High | Legal accuracy — wrong answers have compliance consequences |
| Onboarding Agent | Getting started, tools, processes | Low | Accuracy — outdated content causes friction but limited harm |
All instances share the same base behaviour: - Decline to speculate beyond the library - Always cite sources - Never claim to provide legal, medical, or professional advice - Escalate questions involving individual circumstances that require human judgement
7. Configured Instances¶
7a. Coach Agent¶
Persona name: Configurable (e.g. "Your Coach", "Ember", "The Studio")
Domain: Personal development, performance, wellbeing, goal-setting, skill-building — within the framework of content the deploying organisation has loaded into its library.
Typical library content: Coaching methodologies (e.g. GROW model, solution-focused frameworks), wellbeing resources, learning frameworks, goal-setting guides, curated articles, and any organisation-specific development content.
Platform library: Thinklio provides a platform-level library of general coaching methodology that any Coach Agent can draw from without org-level seeding. Organisations can supplement or replace this with their own content.
Audience: Individual users, typically in a 1:1 private context. Conversations are private to the user — not visible to managers or HR.
Key behavioural differences from base: - More conversational, less formal than HR or Onboarding - Allowed to ask reflective questions and hold a multi-turn coaching dialogue, not just answer questions - Must include a disclaimer when discussing mental health topics directing users to professional support - User knowledge layer is more active — the agent remembers goals, progress, and past conversations to provide continuity
Privacy note: Because coaching conversations are sensitive, the Coach Agent's conversation history and extracted knowledge facts are strictly user-scoped. They are never accessible to org admins, even for audit purposes, unless the user explicitly shares a session.
7b. HR Agent¶
Persona name: Configurable (e.g. "HR Advisor", "People & Culture", "PolicyBot")
Domain: Employment policies, entitlements, procedures, workplace conduct, and compliance — strictly within the content of the deploying organisation's HR library.
Typical library content: Employment contracts templates, leave policies, performance management procedures, workplace health and safety guides, code of conduct, enterprise agreement (if applicable), and any other policy documents relevant to the workforce.
Platform library: No platform-level library for HR — all content must be org-supplied, because policies are inherently organisation-specific and jurisdiction-specific. Using generic HR content would be actively harmful.
Audience: All staff. Tone adapts to query type — direct manager questions get more detailed responses than casual staff questions about entitlements.
Key behavioural differences from base: - Higher confidence threshold — the HR Agent will decline rather than approximate when policy language is ambiguous - Jurisdiction awareness — if the library includes policies for multiple jurisdictions, the agent should ask which applies before answering - Mandatory disclaimer on legal questions: "This information is based on our current policies. For individual advice, please contact HR directly." - All HR Agent conversations are loggable at org admin discretion (unlike Coach Agent). Logging policy must be disclosed to users. - Change detection: when an HR document is updated in the library, the agent flags that its knowledge has changed and prompts users who recently asked related questions to re-check
7c. Onboarding Agent¶
Persona name: Configurable (e.g. "Welcome Guide", "Your Onboarding Buddy", "Getting Started")
Domain: The first-days and first-weeks experience — tools setup, team introductions, process orientation, and practical getting-started guidance.
Typical library content: Onboarding checklists, IT setup guides, systems access procedures, team structure overviews, glossary of internal terms, "how we work" cultural documentation, and any process documentation relevant to new starters.
Platform library: Thinklio provides a platform-level library covering general workplace orientation themes. Orgs supplement with their own onboarding content.
Audience: New employees in their first days, weeks, or months. Tone is warm, patient, and assumes no prior knowledge.
Key behavioural differences from base: - Proactive mode: the Onboarding Agent can be configured to push content (e.g. "Day 3 checklist: have you completed your IT security training?") rather than waiting to be asked - Checklist tracking: integrates with Taskmaster to create and track onboarding task lists - Introductions: can introduce the new employee to relevant contacts by surfacing Rolodex entries for their team members - Time-bounded relevance: content is tagged by onboarding week; the agent prioritises week-relevant content in early conversations - Handoff: after a configured period (e.g. 90 days), the agent proactively transitions the user to the broader knowledge base
8. Use Cases¶
UC-1: Policy question (HR Agent)¶
A new employee asks "how much annual leave do I get in my first year?" The HR Agent retrieves the relevant section from the leave policy document, provides the entitlement figure with the specific policy clause cited, and notes that pro-rata rules apply if they started mid-year, with a reference to the section explaining that.
UC-2: Goal setting (Coach Agent)¶
A user opens the Coach Agent and says "I want to get better at presenting to senior stakeholders." The agent asks a few reflective questions about current challenges, then retrieves relevant content from the library (a presentation skills framework), summarises the key principles, and helps the user set a specific practice goal. The goal is saved to the user's knowledge layer for continuity.
UC-3: Day one (Onboarding Agent)¶
A new starter opens the Onboarding Agent on their first morning. The agent greets them by name, confirms today's date relative to their start date, and surfaces the Day 1 checklist from the library. As they complete items, it updates their Taskmaster task list. It proactively mentions the IT setup guide when they haven't checked that item by midday.
UC-4: Out-of-scope escalation (HR Agent)¶
An employee asks "can my manager legally reduce my hours without notice?" The HR Agent retrieves relevant policy content but recognises this involves individual legal rights that go beyond what a policy document can determine. It provides the policy context it has, but explicitly flags that this requires individual advice and produces a structured escalation note the user can send to HR directly.
9. Open Questions¶
- Should the Coach Agent's strict privacy model (no admin access to conversations) be a hard platform constraint, or a configurable org setting? Some organisations may have legitimate reasons to audit coaching conversations (e.g. compliance in regulated industries). The current position is hard privacy as default, with org-level opt-in disclosure to users that sessions may be reviewed.
- The HR Agent change detection feature (notifying users when a policy they recently asked about has changed) requires the agent to maintain a record of what users have asked about previously. This is a form of user knowledge extraction — is it in scope for the base Persona Agent, or is it HR-specific?
- Multiple library support (an agent drawing from both a platform library and an org library) requires a precedence model: when the org library and platform library give different answers, which takes priority? Default: org library takes precedence, as it reflects the actual deploying organisation's position.
Next: Knowledge Base Agent