Skip to content

Scribe Agent & Digest Agent

Thinklio Built-in Agent Specification — New Additions Version 0.1 | March 2026

These two agents are not in the initial starter catalogue but are identified here as natural extensions of the agent set. Both address gaps that become apparent once the coordinator agents are considered.

1. Scribe Agent

1.1 Purpose

The Scribe Agent operates during meetings and conversations to produce a faithful, structured record. It handles transcription, extracts key information, and produces outputs that other agents (particularly the Meeting Agent and Taskmaster) can act on immediately.

It solves a persistent problem: the gap between what happens in a meeting and what gets recorded. Action items get forgotten. Decisions aren't documented. Context that seemed obvious in the room is invisible to people who weren't there. The Scribe Agent makes the meeting's output as structured and durable as anything produced before or after it.

1.2 Input Modes

The Scribe Agent accepts meeting content in several forms:

Input mode Description
Live audio Real-time audio stream from a meeting or call (via browser integration or telephony connector)
Audio file Uploaded recording (MP3, M4A, WAV, etc.)
Video file Uploaded video with audio track
Third-party transcript Import from Zoom, Teams, Google Meet, or similar (structured or plain text)
Manual notes User pastes or types unstructured notes; Scribe structures them
Integration Direct connection to video conferencing platform APIs

Not all input modes will be available at launch. Audio file upload and third-party transcript import are the lowest-barrier starting points. Live audio requires platform-level audio stream handling.

1.3 UI Structure

[ Chat ]  [ Sessions ]  [ Outputs ]

Chat Tab: Interact with the Scribe Agent about a specific session: - "Summarise the key decisions from yesterday's board meeting" - "What action items came out of the planning session?" - "Who was responsible for the budget review task?" - "Generate a follow-up email for the client meeting"

Sessions Tab: Library of all recorded and processed sessions, searchable by date, participants, and topics. Each session shows its processing status and available outputs.

Outputs Tab: All structured outputs produced from sessions: summaries, transcripts, action item lists, and follow-up drafts.

1.4 Processing Pipeline

For each session, the Scribe Agent produces outputs in stages:

Input (audio / transcript / notes)
    ↓
Stage 1: Raw transcript (if audio input)
    Speaker diarisation (who is speaking)
    Timestamp alignment
    ↓
Stage 2: Structured transcript
    Clean formatting
    Speaker labels resolved to names (from participant list)
    Timestamps normalised
    ↓
Stage 3: Intelligent extraction
    Meeting summary (key topics, key points)
    Decisions made (explicit and implied)
    Action items (what, who, by when)
    Open questions (raised but not resolved)
    Key participants and their positions
    ↓
Stage 4: Downstream outputs
    Formatted meeting notes (markdown)
    Action items → Taskmaster (for each named assignee)
    Meeting log → Rolodex (Interaction record for all participants)
    Follow-up draft → Writer Agent (optional, on request)

1.5 Action Item Extraction

Action items are extracted as structured objects:

ActionItem
├── description     string
├── assigned_to     string (name from transcript)
├── person_id       UUID | null (Rolodex match)
├── due_date        date | null (inferred from transcript language)
├── source_quote    string (the words in the transcript that indicate this action)
└── confidence      float

Low-confidence action items are flagged for human review rather than automatically created as tasks. The user reviews and confirms before Taskmaster records are created.

1.6 Configuration

Setting Description
Default participant list Participants to use for speaker resolution when not specified at session start
Action item confidence threshold Below this, action items are flagged for review rather than auto-created
Auto-create tasks Whether confirmed action items are automatically sent to Taskmaster
Auto-log interactions Whether completed sessions are automatically logged in Rolodex
Transcript storage How long raw transcripts are retained (privacy consideration)
Speaker diarisation Whether to attempt speaker identification in audio input

1.7 Privacy Considerations

Meeting transcription is sensitive. The Scribe Agent must handle participant data with care:

  • Participants should be aware their voice is being recorded and processed
  • Transcripts containing sensitive business information should have appropriate access controls
  • Raw audio is not stored beyond the transcription step unless explicitly configured
  • Transcript content is scoped to the user or team that initiated the session
  • The agent will not share transcript content with agents operating in different scope contexts

1.8 Use Cases

UC-1: Post-meeting processing A user finishes a 60-minute strategy meeting. They upload the Zoom recording to the Scribe Agent. The agent transcribes it, identifies four speakers, extracts eight action items, three decisions, and two open questions. It produces a meeting summary and flags two action items with low confidence for review. The user confirms six action items → Taskmaster creates tasks. The meeting is logged in Rolodex against all four participants.

UC-2: Notes structuring A user has disorganised notes from a call: "spoke to James - he wants revised scope by Friday - budget still tbc - Sarah to check with finance - meeting again next thursday." They paste this into the Scribe Agent. The agent structures it: summary (one paragraph), two action items (James: revised scope by Friday; Sarah: check budget with finance), one open question (budget), one scheduled follow-up (next Thursday). Taskmaster tasks created automatically.

UC-3: Meeting Agent integration The Meeting Agent invokes the Scribe Agent post-meeting as part of its standard post-meeting workflow. The Scribe Agent processes the transcript, returns structured outputs to the Meeting Agent, which uses them to: create Taskmaster tasks, draft a follow-up email for approval, and log the Rolodex Interaction.

2. Digest Agent

2.1 Purpose

The Digest Agent is a scheduled monitor and synthesiser. It watches multiple information sources across a configured time window, filters for what's relevant, and produces a unified briefing delivered to the user on a schedule they control.

It is different from the Personal Assistant's morning briefing in scope and intent. The PA's briefing is about today — tasks, calendar, urgent messages. The Digest Agent's briefing is about the world beyond the user's immediate task list: news on topics they follow, updates from monitored sources, signals from their external environment.

It answers the question: "What do I need to know that I didn't go looking for?"

2.2 Source Types

The Digest Agent can monitor and draw from:

Source type Examples
News searches Topics, keywords, organisations, people
Research feeds Academic publications, preprint servers, on a topic
Web monitoring Specific sites or pages for changes
Internal sources Items in the Enquiry Agent queue, Taskmaster patterns, Rolodex interactions
Agent outputs Completed Research Agent runs, new Knowledge Base additions
Scheduled Research Agent A Research Agent instance running on the same schedule feeds directly into the Digest

2.3 UI Structure

[ Chat ]  [ Digests ]  [ Sources ]  [ Schedule ]

Chat Tab: Query the digest or configure it conversationally: - "Add a digest topic: Australian aged care policy" - "What came through on the Thinklio competitive landscape this week?" - "Change my digest to weekly on Monday mornings"

Digests Tab: Archive of all produced digests, browsable and searchable. Each digest links to its source items.

Sources Tab: Configuration of monitored sources. Each source shows: type, query or URL, last checked, and items found. Add, edit, and remove sources from here.

Schedule Tab: When the digest runs, what format it takes, where it is delivered (in-platform, email, or both), and notification preferences.

2.4 Digest Structure

A digest is a structured document, not a raw feed:

Digest
├── period          string ("Week of 17–23 March 2026")
├── theme_summary   string (one-paragraph synthesis of the most significant developments)
├── sections[]
│   ├── topic       string
│   ├── items[]
│   │   ├── title   string
│   │   ├── source  string
│   │   ├── summary string (2–3 sentences)
│   │   ├── url     string
│   │   └── date    date
│   └── synthesis   string (what the agent thinks is significant about this cluster)
└── flagged_items[] (items the agent has identified as particularly significant)

The agent does not just aggregate — it synthesises. The theme_summary and per-section synthesis fields contain the agent's interpretation of what the collected items mean, why they matter, and what connections it notices between them. Users can configure how much synthesis they want (from "just the links" to "full analysis").

2.5 Configuration

Setting Description
Topics List of topics, keywords, or entities to monitor
Sources Specific sources to include (in addition to general search)
Period Lookback window for each digest (e.g. last 7 days)
Schedule When the digest is produced (daily, weekly, custom)
Delivery In-platform only, email, or both
Synthesis level links_only, summary, full_analysis
Minimum items Don't produce a digest if fewer than N items found (avoids empty digests)
Item limit Maximum items per topic

2.6 Relationship to Research Agent

The Digest Agent uses the Research Agent as a delegate for source retrieval. When the digest schedule fires:

  1. The Digest Agent reads its configured topics and sources
  2. It invokes Research Agent instances (one or more, possibly in parallel) with appropriate parameters
  3. It receives source lists back
  4. It filters, clusters, and synthesises
  5. It produces the structured Digest output
  6. It delivers according to the configured schedule

The Research Agent's update frequency feature (scheduling re-runs) can be used to pre-populate fresh source lists that the Digest Agent then draws from, rather than the Digest Agent triggering Research Agent runs inline.

2.7 Use Cases

UC-1: Weekly professional digest A user configures a weekly digest covering: Australian aged care regulation, AI in healthcare, and Perth commercial property. Every Monday morning they receive a structured briefing covering developments across all three topics from the previous week, with the Digest Agent's synthesis of what's significant.

UC-2: Competitive monitoring A business development manager configures a daily digest monitoring competitors and industry news. The digest flags when a competitor makes a significant announcement, surfaces relevant analyst commentary, and notes any client-relevant developments. It arrives in their inbox each morning before they start the day.

UC-3: Internal digest A team lead configures a digest that monitors internal sources: new Items in the support queue (volume and type), completed Research Agent runs (what new research has been produced), and Taskmaster patterns (overdue task trends). This gives them a weekly operational pulse without having to check each system manually.

3. Open Questions

Scribe Agent

  • Live audio transcription requires real-time streaming infrastructure that is qualitatively different from file upload. Should live audio be deferred to a later phase, with file upload and transcript import as the initial scope?
  • Speaker diarisation quality varies significantly by audio quality and number of speakers. How does the agent handle poor audio? A graceful fallback to "Speaker 1, Speaker 2" with user-resolved names is better than failing.
  • Retention of raw audio and transcripts is a significant privacy and compliance consideration. Is a configurable retention window sufficient, or should raw audio deletion be mandatory and immediate after transcription?

Digest Agent

  • The synthesis feature requires the agent to make editorial judgements about what matters. This is valuable but introduces potential for bias or missed significance. Should synthesis be opt-in, with raw aggregation as the default?
  • Digest delivery via email means the Mail Agent is in the dependency chain for the Digest Agent. The Digest Agent should degrade gracefully if the Mail Agent is not configured — deliver in-platform only.
  • How should the Digest Agent handle a week with very little matching content? A "nothing significant this week" digest is still useful. An empty digest is not. The minimum_items config setting addresses this but needs a good default.

These agents are proposed additions to the starter catalogue. They should be reviewed for inclusion in the Phase 2 agent set.