Thinklio Agent Backlog & Future Ideas¶

Version: 0.2 | March 2026 Status: Living document — update as gaps emerge from usage Audience: Product, development

Purpose¶

This document captures agent ideas, capability extensions, and platform concepts that are not in the current built-in agent set. It is not a roadmap — it does not imply commitment or priority beyond the flags noted. It is a backlog: a place to hold ideas so they are not lost, reviewed when planning, and updated as the platform gets used and real gaps emerge.

The most valuable entries in this document will probably be ones added after the platform is running. Theoretical gaps are speculative. Usage gaps are real.

Priority flags:

🔴 Soon — addresses a gap in the current set that will be felt early; consider for the next planning cycle
🟡 Medium — valuable but not urgent; natural phase 2 or 3 addition
⚪ Later — good idea, low urgency, or dependent on platform maturity
💡 Custom — probably best built as a custom agent in Agent Studio rather than a built-in

Section 1: New Built-in Agent Proposals¶

Note: The CEO Agent (spec 25) and Board Agents — Board Member Agent and Board Secretariat Agent (spec 26) — have been fully specced and are candidates for the built-in catalogue. They are not listed here as they already have dedicated spec documents. They should be formally added to the starter catalogue once reviewed.

1.1 Clip Agent 🔴¶

What it does: Lightweight content capture and save-for-later. Users save URLs, paste text, or forward content to Clip; the agent fetches and stores it, generates a summary, auto-tags it, and makes it searchable. Saved items can be forwarded to the Research Agent as source material, added to the Knowledge Base, or converted to Taskmaster tasks.

Why soon: High daily utility, small scope, fills an obvious gap in the personal productivity cluster. Users will immediately want somewhere to put things they find while browsing or reading. Without it, they resort to browser bookmarks or email-to-self, neither of which connects to the rest of the platform.

Rough scope: URL fetching, content extraction, summarisation (delegates to existing pipeline), tagging, search. No new infrastructure required — uses the existing media and library system.

Related agents: Research Agent (source input), Knowledge Base Agent (archive destination), Taskmaster (action conversion).

1.2 Contracts Agent 🔴¶

What it does: Contract lifecycle management. Drafts contracts from templates, tracks negotiation state, flags clauses for review, monitors execution and signature status, tracks renewal and expiry dates, and alerts on upcoming obligations.

Why soon: High value for any business user. Organisations create, negotiate, and manage contracts constantly. The gap is felt immediately once Thinklio is used for business operations. Contract drafting from templates is close to what the Writer Agent does, but the lifecycle tracking (status, obligations, renewals, counterparty) requires dedicated data objects and ongoing monitoring.

Rough scope: - Contract templates (stored in the library system) - Contract record (parties, status, key dates, obligations) - Drafting (delegates to Writer Agent with contract-specific templates) - Review assistance (clause analysis, flagging unusual terms) - Renewal and expiry monitoring (delegates to Monitor Agent or runs as scheduled check) - Signature status tracking (integration with DocuSign, Adobe Sign, or similar)

New data objects needed: Contract, ContractParty, ContractObligation, ContractEvent.

Related agents: Writer Agent (drafting), Fact Checker Agent (clause verification), Monitor Agent (deadline alerting), Rolodex (counterparty records).

1.3 Webhook / Event Agent 🔴¶

What it does: Receives inbound structured events from external systems via webhook endpoints and routes them into the Thinklio platform as Items, notifications, tasks, or agent triggers. Each configured webhook endpoint has a processing rule: what to create, who to notify, which agent to invoke.

Why soon: Completes the inbound channel picture. Mail and Messenger handle human-initiated inbound contacts. This handles system-initiated inbound events: a new GitHub issue, a Stripe payment failure, a form submission from a website, a deployment alert from a CI/CD pipeline. Without this, integrating external systems requires polling or custom code.

Rough scope: - Webhook endpoint registration (unique URL per configured integration) - Payload schema mapping (user defines how fields map to Thinklio objects) - Processing rules (if payload matches condition → create Item / create Task / invoke agent / notify user) - Delivery confirmation and retry logic - Payload log for debugging

Note: This is closely related to the Integration API described in the PRD. The Webhook Agent is the user-facing product layer on top of that infrastructure — the admin UI for configuring what happens when an event arrives, rather than a raw API capability.

Related agents: Enquiry Agent (Item creation), Taskmaster (task creation), Monitor Agent (condition-triggered actions).

1.4 Translation Agent 🟡¶

What it does: Translates content between languages with terminology consistency, brand voice preservation, and glossary management. Unlike the ad-hoc translation capability in the Chat Agent or Writer Agent, this maintains a translation memory and organisation glossary so terms are translated consistently across all content.

Why medium: The platform's internationalisation architecture targets nine languages. As Thinklio is used across language boundaries, translation quality and consistency become product differentiators. The gap is real but not felt until the platform has multilingual users or content.

Rough scope: - Translation from any language to any configured target language - Organisation glossary (specific terms that must be translated a certain way) - Translation memory (previously translated segments for consistency) - Brand voice preservation (same voice profile in the target language) - Batch translation (translate a full document, not just a passage)

Related agents: Writer Agent (content input), Content Agent (multilingual content production), Report Writer Agent (multilingual PDF output).

1.5 Health Agent 🟡¶

What it does: Monitors the Thinklio platform itself — agent performance, error rates, credit consumption, unusual activity patterns, and system health — and reports to workspace admins. Distinct from the Monitor Agent, which watches external systems and business conditions.

Why medium: As the platform grows and more agents are deployed, admins need visibility into how the system is performing without becoming infrastructure engineers. The Health Agent translates platform telemetry into actionable admin intelligence: "The Research Agent is consuming 3x its normal credits this week — three jobs failed and were retried multiple times" is more useful than a raw metrics dashboard.

Rough scope: - Credit consumption tracking and anomaly detection - Agent error rate monitoring - Job failure patterns and retry analysis - Usage trends per agent, team, and user - Weekly health digest to workspace admins - Alert on critical thresholds (agent down, budget approaching limit, unusual access patterns)

Related agents: Monitor Agent (shares alerting patterns), Finance Agent (credit/cost context).

1.6 Forms Agent 🟡¶

What it does: Creates, distributes, and processes structured forms — surveys, intake questionnaires, feedback forms, approval requests. Responses are collected into structured records, summarised by the agent, and routed to the appropriate place (Enquiry Agent Items, Taskmaster tasks, Knowledge Base entries, or direct to the requesting user).

Why medium: Forms are a universal business tool. Currently the Enquiry Agent handles unstructured inbound contact. Forms handle structured inbound data. The combination covers almost all intake scenarios.

Rough scope: - Form builder (question types: text, multiple choice, rating, date, file upload) - Distribution (link, email via Mail Agent, embedded widget) - Response collection and storage - AI summarisation of open-text responses - Routing rules (where do responses go?) - Repeat / scheduled forms (e.g. weekly team check-in)

Related agents: Enquiry Agent (structured inbound), Taskmaster (action items from responses), Digest Agent (periodic response summaries).

1.7 Delegation Tracker 🟡¶

What it does: Tracks tasks and decisions delegated to other people — not agents, but human colleagues and external parties. "I asked three people to do things; what's the status of each?" Monitors waiting items, sends follow-ups when things are overdue, and keeps a clear record of what has been delegated, to whom, and by when.

Why medium: The Taskmaster handles tasks assigned to the user. The Project Coordinator handles tasks within a project. Neither cleanly handles the personal delegation pattern: things the user has asked others to do and needs to follow up on. This is a very common source of dropped balls.

Rough scope: - Delegation record (what, to whom, by when, current status) - Overdue detection and follow-up prompting - Integration with Mail Agent (send a follow-up via email) - Integration with Pulse (send a follow-up message) - Weekly delegation summary ("you're waiting on 5 things from others")

Possible approach: Could be implemented as a Taskmaster extension (a "waiting on" task type) rather than a separate agent. Worth evaluating before building standalone.

Related agents: Taskmaster (waiting tasks), Mail Agent (follow-up emails), Pulse (follow-up messages), Rolodex (counterparty context).

1.8 Expense Receipt Agent ⚪¶

What it does: Reads expense receipts (photos, PDFs, email forwards) and automatically creates expense records in the Finance Agent. Uses OCR and extraction to pull merchant, amount, date, and category from the receipt without manual entry.

Why later: This is a natural extension of the Finance Agent rather than a standalone agent. The media processing pipeline (via the existing document ingestion system) already handles the OCR and extraction part. The Finance Agent just needs a "process receipt" capability added to it. Flag here for completeness but implement as a Finance Agent feature.

1.9 Learning Agent ⚪¶

What it does: A structured learning companion. The user sets a learning goal (a skill, a topic, a certification) and the agent builds a personalised learning path, surfaces relevant resources from the Knowledge Base and web, tracks progress, tests understanding through questions, and adapts the path based on what the user is finding easy or hard.

Why later: High value but significant scope. The Coach Agent (in its configured form) covers some of this ground. The distinction is structure — the Learning Agent is specifically about guided, tracked skill acquisition rather than general coaching. Requires a learning path data model and progress tracking that goes beyond the current knowledge layer architecture.

Related agents: Coach Agent (overlapping domain), Knowledge Base Agent (content source), Research Agent (external resource discovery), Taskmaster (study schedule).

1.10 Announcement Agent ⚪¶

What it does: Manages internal announcements and broadcasts. An org admin or team lead composes an announcement; the agent distributes it via configured channels (Pulse, email, Messenger), tracks who has seen it, sends reminders to those who haven't, and archives announcements with a searchable history.

Why later: The Pulse broadcast thread type covers the basic case. A dedicated Announcement Agent adds read-tracking and multi-channel distribution, which matters at org scale but is not needed early.

Section 2: Extensions to Existing Agents¶

These are capabilities that belong in an existing agent rather than a new one, but are worth tracking explicitly.

2.1 Calendar Agent — Booking Links 🔴¶

What it adds: Outbound scheduling links (Calendly-equivalent). The user generates a booking link that shares their availability, the recipient picks a time, and the event is created automatically without back-and-forth. Essential for customer-facing scheduling.

Why soon: This is one of the most common use cases for calendar tooling in a business context. Without it, external scheduling still requires the email ping-pong that Thinklio is supposed to eliminate.

2.2 Research Agent — Citation Styles 🟡¶

What it adds: Full support for academic citation styles (APA, Chicago, Vancouver, Harvard, IEEE) in source list output, not just hyperlinks. Currently the Data Agent handles citation list export but the format options are limited.

Why medium: Important for academic, clinical, and professional contexts. The Clindice use case (clinical reference) makes this particularly relevant for Novansa's own products.

2.3 Mail Agent — Attachment Intelligence 🟡¶

What it adds: Reading, summarising, and routing email attachments. A PDF invoice attached to an email is detected, extracted, and forwarded to the Finance Agent. A contract attachment is detected and forwarded to the Contracts Agent. A document attachment is summarised inline in the mail view.

Why medium: Email attachments are where a significant proportion of business-critical information lives. The Mail Agent currently treats them as opaque. Adding attachment intelligence makes the Mail Agent substantially more useful for business workflows.

What it adds: An embeddable web form widget that creates Enquiry Agent Items directly from a website. A visitor fills in a contact form; the submission arrives as an Item without going through email.

Why medium: This was flagged as deferred in the Enquiry Agent spec. It requires a webhook endpoint (see Webhook Agent above) and a small embeddable widget. High value for any customer-facing deployment.

2.5 Scribe Agent — Live Audio 🟡¶

What it adds: Real-time audio transcription during a live meeting, rather than file upload after the fact. Requires streaming audio from the browser or a phone integration.

Why medium: File upload and transcript import cover most use cases for now. Live audio is the premium capability and requires significantly more infrastructure (real-time audio streaming, speaker diarisation at speed). Right call to defer but worth planning for.

2.6 Knowledge Base Agent — User Contributions 🟡¶

What it adds: A formal contribution workflow where users can propose documents or facts for the knowledge base, which an admin approves before indexing. Currently the knowledge base is admin-only.

Why medium: Institutional knowledge lives with the people doing the work, not just the people managing the systems. A contribution workflow surfaces that knowledge without opening the knowledge base to uncontrolled edits.

2.7 Visualiser Agent — Interactive Charts ⚪¶

What it adds: Filterable, drill-down, and animated chart outputs using the Vega-Lite interactivity model. Currently all Visualiser Agent outputs are static renders.

Why later: Static charts cover the vast majority of use cases. Interactive charts require a client-side execution model that adds platform scope. Phase 2 of the Visualiser Agent as originally planned.

2.8 Pulse — Unified Attention Queue ⚪¶

What it adds: A single attention queue spanning both Pulse (internal) and the Messenger Agent (external platforms), scored together so the most important message surfaces regardless of whether it came from a colleague in Pulse or a client on WhatsApp.

Why later: Powerful but requires the two agents to share a scoring model and a unified UI. The individual agents need to be stable first. This is the natural integration once both are running.

2.9 Finance Agent — Budget Management 🟡¶

What it adds: Proactive budget tracking with department/team breakdowns, spend forecasting, and variance alerts. Currently the Finance Agent handles expense submission and policy queries. Budget management adds the forward-looking dimension.

Why medium: Once the Finance Agent is deployed and expense data is accumulating, the next natural question is "how are we tracking against budget?" The data is there; it just needs a reporting layer.

Section 3: Platform Capabilities (Not Agent-Specific)¶

These are platform-level capabilities that would make multiple agents more powerful, rather than new agents in their own right.

3.1 Agent Marketplace 🟡¶

What it is: A catalogue of community-built and partner-built agent templates, available to all workspaces. Users can browse, install, and deploy templates built by others. Thinklio curates quality and safety.

Why medium: The Agent Studio makes it easy to build custom agents. The marketplace makes it easy to share them. This is the network effect mechanism for the platform — as the community builds useful agents, every workspace benefits.

3.2 Agent Analytics Dashboard 🟡¶

What it is: A workspace-level view of agent usage, performance, cost, and value. Which agents are used most? Which cost the most? Which have the highest user satisfaction? Which are underused relative to their cost?

Why medium: Admins need this to manage their deployment. The Health Agent covers the technical dimension; this covers the business value dimension.

3.3 Cross-Agent Search 🟡¶

What it is: A unified search across all content in the workspace — knowledge facts, media library, Pulse threads, Enquiry Agent Items, Taskmaster tasks, and agent-produced artefacts. Currently each agent has its own search. A single search bar that spans everything is a significant usability improvement.

Why medium: Users should not need to know which agent holds the information they are looking for. They should just be able to search and find it. The underlying data is accessible; this is a product layer over the top.

3.4 Scheduled Agent Runs 🟡¶

What it is: The ability to schedule any agent (not just the Research Agent, which already supports this) to run on a recurring basis. A Digest Agent runs weekly. A Finance Agent produces a monthly report. A Health Agent sends a weekly admin briefing. Currently scheduling is agent-specific; this makes it a platform capability available to all agents.

Why medium: Recurring automation is a core use case for a platform like Thinklio. Making it universal rather than per-agent is the right architectural direction.

3.5 Agent Handoff Protocol 🟡¶

What it is: A standardised protocol for one agent to hand off a conversation or task to another agent with full context preserved. Currently agents are invoked by coordinators, but there is no clean pattern for a specialist agent to recognise it has hit the edge of its domain and hand off gracefully to another agent or to a human.

Why medium: "I can't help with that, but X agent can, and here is the context it needs" is a powerful capability. Without a handoff protocol, users hit dead ends or have to re-explain their context to a new agent from scratch.

3.6 Agent Versioning ⚪¶

What it is: The ability to version agent configurations, roll back to a previous version, and run A/B comparisons between versions. When an agent's system prompt or knowledge is updated, the old version is preserved.

Why later: Important for mature deployments where agents are business-critical. Premature for early-stage usage. Implement when the platform has customers who would be seriously affected by an agent configuration regression.

3.7 Agent Skills Layer 🔴¶

What it is: Named, parameterised, output-typed sub-capabilities that live inside an agent and can be invoked explicitly by users or coordinating agents. A skill is more than a tool call (it involves LLM reasoning and prompt engineering) but less than a full agent delegation (no separate context assembly, knowledge layers, or job tracking). It fills the gap between deterministic tool calls and heavyweight agent delegation.

Why soon: Skills should inform the Agent Studio design from the start — retrofitting them later is harder than designing for them from the beginning. They are particularly important for coordinator agents, where consistent, typed outputs from specialist delegates determine the quality of assembled results. See the Agent Studio spec (section 3) for the full skills model.

The three-tier model this completes: - Tools — deterministic, code-based operations (calendar lookup, API call) - Skills — prompt-based, reusable operations with typed I/O (summarise, extract, classify) - Agents — domain work with persistent state and multi-step reasoning

Trigger for implementation: When three or more agents share the same prompt-based operation and the duplication starts to hurt — or when the Agent Studio is being designed, whichever comes first.

Examples of platform skills already identified: - Summarisation Agent: summarise.executive, summarise.bullet, summarise.quiz, summarise.actions - Writer Agent: write.headline, write.social, write.expand, write.condense - Research Agent: research.quick, research.deep, research.cite - Fact Checker: check.factual, check.voice, check.full - Scribe Agent: scribe.summary, scribe.actions, scribe.decisions - Data Agent: data.to_csv, data.top_n, data.diff

Skill scopes: Platform (built by Thinklio), org (defined by workspace admins), user (personal shortcuts on personal agents).

Open questions: - Should skills have a lighter cost model than full agent interactions? - Can coordinator agents generate skill invocations dynamically at runtime, or are skills always pre-defined? - Should org-defined skills be shareable between workspaces via the template marketplace?

What it is: Voice input to agents (speech-to-text), image input beyond document analysis (e.g. photographing a whiteboard and having the agent extract and structure the content), and video analysis. Currently the platform handles text and documents.

Why later: High value eventually, significant infrastructure scope. The Scribe Agent's live audio transcription is the natural entry point for voice. Image input is closer — the media processing pipeline already handles image analysis. A coherent multi-modal strategy is a later-phase product decision.

Section 4: Ideas From Usage (To Be Filled In)¶

This section is intentionally empty. It will be populated as the platform is used and real gaps emerge. Theoretical gaps are speculative. Usage gaps are real.

When adding an entry here, note: what were you trying to do, what could you not do, and what would have helped?

Date	Context	Gap observed	Suggested solution	Priority

Section 5: Ideas Deliberately Out of Scope¶

These are ideas that came up during spec development and were consciously set aside. Documented here so the decision is not re-litigated without new information.

Idea	Why out of scope
Image generation agent	Complexity, cost, and rapidly evolving external tooling. Better to integrate a best-in-class external service via the Integration API than build natively.
Accounting / bookkeeping	The Finance Agent handles expense management and financial queries. Full accounting (general ledger, reconciliation, tax) is a specialist domain with regulatory implications. Out of scope for Thinklio; integrate with Xero/QuickBooks instead.
CRM system	Rolodex provides the relationship layer. A full CRM (pipeline management, forecasting, territory management) is a different product category. The Customer Intelligence Agent bridges the gap for most use cases.
Project management system	The Project Coordinator and Taskmaster cover project coordination. A full PM system (Gantt, resource management, earned value analysis) is a different product category. Integrate with Linear/Jira/Asana for teams that need this depth.
Video conferencing	Integrate with Zoom/Teams/Meet rather than build. The Scribe Agent handles the meeting intelligence layer.
Social media publishing	The Content Agent produces content. Publishing to social platforms is an integration, not a Thinklio feature. Use Buffer/Hootsuite/native APIs via the Integration API.

This document should be reviewed at the start of each planning cycle and updated after any significant usage period. The most important updates will come from the team using the platform, not from theoretical planning.