Digital Worker Strategy

A digital employee that can use your tools, retain working knowledge, and show how it reached an answer.

Digital Worker Strategy adds Agentron's digital employee runtime to EmploidAI. Give a worker a role, connect its model and permitted tools, and supply the knowledge it should use. It can carry on persistent conversations, execute Agent Flows, work with files, maintain structured knowledge, and run scheduled work. Its Brain and Knowledge interfaces make that state visible inside the worker's sidebar.

The plugin is included with an EmploidAI subscription. Model usage, third-party services, and infrastructure can incur separate charges.

What you can do

• Ask questions grounded in your worker's job definitions, wiki documents, and accumulated knowledge.
• Run existing Agent Flows and use connected tools within the worker's configured permissions.
• Turn connected email and meeting archives into source-linked notes about people, organizations, projects, and topics.
• Inspect a knowledge graph, follow links back to notes, and review which documents a response actually used.
• Use Brain to identify conflicting or stale information, possible duplicates, and useful follow-up questions.
• Keep conversations, workspace files, task state, and knowledge available across worker restarts when persistent storage is retained.
• Review conversation and model-usage history through the host platform's Interactions interface when that service is available.
• Manage reusable commands, automations, to-dos, and background jobs from the worker's own interfaces.

The interfaces included with your worker

Interface	What it is for
Brain	Configure reflection schedules and sleep behavior; inspect runs, evidence, proposals, and knowledge corrections.
Knowledge	Organize knowledge scopes and documents, edit Markdown, configure knowledge agents, and import or export knowledge.
Knowledge graph	Explore entities, documents, and their relationships; open the underlying notes behind a connection.
To-dos	Track the worker's work and follow-up items.
Background jobs	Configure scheduled or triggered work and inspect its executions and outputs.
Commands	Maintain reusable instructions and entry points for recurring requests.
Automations	Manage the capabilities and instructions that shape the worker's responsibilities.

These are contributed Worker UIs. They appear for an employee using this strategy on an EmploidAI installation that supports the Worker UI platform.

From source material to useful knowledge

The worker maintains readable Markdown notes alongside indexes and graph data. It separates source archives from the knowledge distilled from them, so you can inspect both the evidence and the resulting note.

The automatic extraction pipeline works with email threads and meeting archives delivered to the worker's workspace by your configured integrations. Installing the strategy alone does not connect a mailbox, join meetings, or grant access to external accounts.

1. Discover changed sources. The knowledge worker tracks source metadata and content hashes so unchanged sources are not extracted again on every scan.
2. Filter and classify. Email classification helps separate durable information from noise. A receipt or an automated notification need not become a permanent memory.
3. Resolve identities. The note-creation agent searches existing notes before adding an entity. Matching names alone are insufficient; email addresses, domains, aliases, or documented context provide stronger evidence.
4. Extract durable information. Notes can capture roles explicitly stated in the source, decisions, commitments, working relationships, open items, and supported changes in status.
5. Update and link. Existing notes receive relevant updates; source references and supported relationships connect the information to its context.
6. Tag and curate. Separate background passes organize note metadata and maintain the knowledge collection. The graph and index are rebuilt from the resulting notes.

What happens when a person is extracted?

Suppose a connected meeting transcript identifies Maya at Northstar as the contact for an ongoing rollout. If the evidence meets the extraction rules, the worker can create or update Maya's People note, connect it to Northstar's Organization note, and record the source-supported activity or commitment. You can then open the note or navigate the relationship in the graph.

This is a knowledge note in the worker's workspace. It does not by itself create a platform user, send Maya a message, or add a contact to an external CRM. Those actions require a separately configured tool or workflow and the applicable authorization.

Extraction follows deliberately conservative rules:

• The memory owner does not become a separate People entity.
• A new note is not created for every name or meeting attendee.
• A purely inbound email does not create new canonical notes; email creation is gated by evidence of the owner's engagement. Supported changes may still update an existing note.
• A relationship must be supported by the source or an existing note. Two names appearing in the same processing batch are not enough.
• New Projects and Topics found by background extraction become suggestions for explicit promotion. Existing canonical Project and Topic notes can be updated.
• Source text is treated as evidence, including when it contains instructions. Extraction tools restrict writes to the knowledge areas intended for notes and suggestions.

Generated notes and links remain reviewable. Model mistakes are possible, so source references matter when the information informs a decision.

How the knowledge graph is structured

The graph represents several kinds of nodes:

• People and Organizations capture durable entities and their context.
• Projects and Topics capture explicitly maintained areas of work or interest.
• Wiki documents, job definitions, and agent notes connect reference material and instructions to the broader knowledge collection.

Edges come from recorded relationships and resolvable links in the underlying documents. The graph is an index of the worker's documented knowledge, not a guarantee that every real-world relationship has been discovered. You can inspect and correct the underlying content when an identity, connection, or fact is wrong.

The Knowledge Tree organizes scopes and documents for authoring; the graph helps you explore connections across them. Promoted projects and topics can also be configured for periodic updates through their live-note settings.

Brain: reflection and review

Brain performs a bounded reflection pass over a snapshot of the worker's knowledge. It looks for possible links or merges, contradictions, stale facts, and unanswered questions. Each proposal carries evidence references, a rationale, and confidence/risk information for review.

You can configure the schedule, timezone, selected days, and limits such as duration, tokens, and proposal count. Brain supports manual wake/sleep controls and two sleep behaviors: soft sleep keeps responding, while strict sleep queues incoming replies for later processing.

A Brain run produces a report and reviewable proposals. It is not model training, and a proposal is not automatically established as a fact. When resolving an issue, you can provide an authoritative clarification, compare the proposed Markdown with the current note, and apply a correction through the review interface. The initial reflection pass does not silently rewrite all source notes.

Completed run results are persisted so retrying the reporting step does not require repeating a completed model call.

Tools, Agent Flows, and repeatable work

The strategy uses the employee's configured model and credentials through EmploidAI's model bridge. Provider and credential availability depend on the host platform and the account connected by your administrator.

Tools are discovered and described through the platform bridge. For an Agent Flow, the worker can find the relevant flow, inspect its inputs, and execute it with the appropriate arguments. Access is governed by the employee's configuration and platform checks; installation is not blanket access to every integration or tenant.

Commands and automations let you define repeatable behaviors and responsibilities. Background jobs support work that should run on a schedule or a configured trigger. Available actions depend on the tools and integrations installed in your workspace.

Traceability and context use

The strategy reports interaction progress, completed messages, and model-usage events to the host platform. Delivery receipts are retained for retries after a reconnect. On a compatible host, its interaction-history service can import existing conversation logs in bounded pages, including available attachment references. The host provides the Interactions interface; historical records can only include data retained in the worker's storage.

A response can expose a Knowledge trace showing loaded knowledge nodes and orchestration steps. The trace can include model calls, measured document context, provider-reported usage, tool schemas, messages, and timing when available. This helps distinguish the amount of knowledge loaded from total model usage across multiple calls.

The runtime includes selective knowledge loading and reusable operation context. New tool results reach the model in full after sensitive-data sanitization, without automatic result compression or hidden record arrays. Artifacts created by earlier versions remain recoverable. Actual usage still depends on conversation history, tool schemas, the selected model, and the number of reasoning and execution steps.

Files, persistence, and data access

The worker runs as an OCI container with persistent workspace and state directories. It can create working files and artifacts and retain knowledge and conversation state across restarts. Retention depends on the host's storage and backup configuration; replacing or deleting those volumes can remove the stored state.

During a task, the worker may process conversation messages, attachments, selected knowledge, tool inputs and outputs, and configured email or meeting archives. Relevant content is sent to the model provider chosen for that employee. Connected tools can send data to their respective services when invoked. Background knowledge extraction and Brain runs can also use the configured model.

The image is pinned by digest and declares a non-root runtime, a read-only root filesystem, writable workspace/state mounts, and no required Docker socket. These container settings complement the platform's authorization checks and your organization's data-handling policies.

See the plugin privacy and data-handling notice for details.

Getting started

1. Install Digital Worker Strategy from the Agentron publisher in your tenant's Marketplace.
2. Create or open an employee and select this strategy.
3. Configure the employee's model and credential, role, job definitions, and allowed tools or Agent Flows.
4. Confirm that the worker starts and that Brain, Knowledge, and the other contributed interfaces appear in its sidebar.
5. Add reference knowledge. Connect email or meeting sources only if you want their archives processed.
6. Begin with a focused request, inspect its response and trace, and review the knowledge notes it uses.
7. Enable Brain schedules, live notes, or background jobs when their scope and model usage suit your needs.

The release includes Linux AMD64 and ARM64 images. The host must support the EmploidAI worker strategy platform, interaction history, sessions, artifacts, knowledge traces, and Worker UI services. A working model credential is required for model-driven tasks. External integrations are configured separately.

Support and release notes

For installation, account, or runtime help, use EmploidAI support. Include the plugin version, affected worker, and a relevant error or trace with confidential data removed. Publishing and platform documentation is available in the Marketplace documentation.

0.55.11: Keeps harness-only receipt metadata out of provider requests while preserving complete tool content. Validated with the opportunity-listing regression on the local Sales Expert worker.

0.55.10: Preserves full governed tool receipts across native execution and replay, applies execution economy without an intent plan, and exposes exact knowledge source paths for direct lookup. Pair with the platform Collection pagination fix for deterministic pages.

0.55.9: Removes preliminary intent planning and lane routing. Conversations and durable object sessions execute directly in the main agent, which selects capabilities through the existing permission and approval boundaries.

0.55.8: Removes automatic tool-result compression and the bundled TinyJuice binary. Preserves complete sanitized results and recovery of artifacts from earlier conversations.

0.55.7: Prepares the current collection skill and live schemas automatically for every relevant turn, including read-to-write follow-ups. Prior injected catalogs are omitted from model context while durable history and authorization remain intact.

0.55.6: Reduces collection prompt/context overhead, prepares live operation schemas during skill loading, and keeps small field schemas visible. Removes the new turn/hour/day spending quotas so ordinary conversations continue. Activates early job authorization checks; confirmed write receipts remain recoverable after an interrupted run.

0.55.5: Preserves routing identities in batched metadata receipts, calibrates per-request reservations and improves budget error messages.

0.55.4: Adds bounded collection counting, compact receipts, consistent result extraction, early job-reference checks, and persistent model usage limits across conversations and background workloads. Requires the accompanying platform MCP/count and usage-accounting update. The default spending guard uses conservative configured rates and is not a provider invoice.

0.55.3: Renames Competencies to Automations in worker navigation, panel headings, and creation labels.

0.55.2: Routes Drive reads through the installed skill and shared MCP catalog. Removes the legacy native read tool and conflicting prompt instructions, so Drive calls use the backend's current required parameters and authorization.

0.55.1: Adds interaction-history export and durable reporting of interaction progress, messages, and model usage to compatible EmploidAI hosts. Reports carry the strategy version for traceability. Existing plugin and strategy identities are retained for upgrades.

0.54.9: Expanded the public guide, documented extraction and Brain behavior, and added signed pricing, permission, support, compatibility, and data-handling disclosures.

0.55.12: Resolves Collection read-filter labels from live scoped metadata; prevents ambiguous substitutions and unnecessary broad reads.

0.55.13: Aligns read-filter guidance with runtime label resolution to avoid redundant schema-only model turns.

0.55.14: Large Drive document pagination and authenticated local file transfer; uploaded PDF/DOCX previews link to complete extracted text. Deploy with the matching platform backend update.

0.55.28: Web searches distinguish empty results from provider blocking. Loop guard guidance is delivered with tool results, preserving completed analyses. This production maintenance release builds on 0.55.14.