Go to the internal Admin interface and select the relevant client or Business Unit.
Approved context and task notes
Context Management.
Know what Jessie treats as approved client context, what she pulls as evidence, and what stays a one-off note from you.
Context Types
| Type | What it means | How Jessie uses it |
|---|---|---|
| Approved client context | Reviewed client facts, rules, strategy, priorities, measurement notes, and source references. | Injected into skill prompts and used to judge performance in context. |
| Retrieved context chunks/facts | Prompt-ready slices from the R2-backed client context lake and D1 context facts. | Loaded at run time through the context retrieval interface. |
| Task-specific consultant notes | Instructions and caveats on the Asana task or follow-up comments. | Guide the current run only; do not become durable truth automatically. |
| Live evidence | Fresh or configured source data from GA4, Search Console, Ads platforms, Browser Run, R2, or D1. | Grounds the report or audit. |
| Generated output | Reports, comments, attachments, evidence JSON, and planning checkpoints. | Reviewed by a human before client use. |
Agent Context Folder
Give every client or Business Unit Jessie works on a maintained Box Agent Context folder. You own it, and it's the source of truth Jessie draws on. What you put here is what she trusts.
| Responsibility | What to do |
|---|---|
| Create the folder | Create or confirm a Box folder named Agent Context inside the client or Business Unit context area. |
| Grant access | A Box admin or folder owner must make the folder accessible to the AgentOS Box service account. The system checks access through the configured Box binding. |
| Keep it curated | Store approved briefs, rules, strategy notes, current priorities, measurement caveats, and other context Jessie should rely on. |
| Use generated indexes as maps only | 00-agent-context-index.md can help humans locate files, but Jessie treats it as locator metadata rather than reviewed client context. |
| Remove or replace stale material | Outdated files should be removed, archived outside the folder, or replaced with clearer current context so future runs do not inherit old assumptions. |
| Promote generated output deliberately | A Jessie report, Asana note, or generated artifact is not trusted context just because it exists. It becomes source context only when a human deliberately adds the approved version to the Agent Context folder. |
Consultants should maintain the Box folder, not D1 rows, R2 objects, chunks, embeddings, or derived facts. Those are system-managed cache/index layers behind the context retrieval interface.
Admin Configuration
Jessie can't see the folder until you link it to the Business Unit in the Admin interface. Until that's done, the files just sit there.
The Context tab owns the Business Unit context source binding.
Paste the Box Agent Context folder URL or numeric folder ID.
The Worker validates and normalizes the folder reference, then stores the Context Source Binding in D1.
Use Test Access to confirm the AgentOS Box service account can list the configured folder.
Run the context refresh so the system scans the folder, records the manifest, downloads supported text files, writes chunks to R2/D1, and creates a prompt-ready Context Pack when usable content exists.
| Admin state | Meaning for Jessie |
|---|---|
| No active Context Source Binding | Context-dependent Jessie work should block or ask for setup instead of producing context-free recommendations. |
| Access test failed | The Box folder exists in Admin, but the AgentOS service account cannot read it yet. Fix Box permissions before relying on it. |
| Refresh completed without prompt-ready chunks | The folder was scanned, but current supported content was not available for prompts. Add supported text/markdown context or wait for parser coverage. |
| Prompt-ready current Context Pack | Jessie can retrieve relevant context chunks/facts from the current pack during task execution. |
Context Flow
Box remains the business-maintained source of truth through the configured Agent Context folder.
The Business Unit Context tab stores the active Box Context Source Binding and verifies service-account access.
The context lake extracts and indexes prompt-ready non-index chunks and facts without giving the model unrestricted memory. Slow live refresh falls back to the current prompt-ready pack rather than leaving a run leased indefinitely.
The task is matched to a configured Business Unit and its source assets.
Approved context, consultant notes, processing-folder signals, communications context, artifact capsules, and evidence pointers are compacted for the selected skill without dumping raw memory into the prompt.
For general-task workbench runs, box.client_context_read reads compact approved context metadata/source pointers from the current context pack or active Context Source Binding without dumping raw chunks.
The model receives bounded context, retrieval policy, and evidence; raw files and long chunks stay on demand.
Reports should separate approved context, task notes, data evidence, and caveats when material.
Agency Brain Lean Context
Agency Brain v1 is the task-time context layer over the existing Business Unit resolver, Context Source Binding, R2-backed context lake, D1 facts, report artifacts, and Asana thread state. It tells Jessie what is known and where to retrieve more, without turning Box, old reports, or communications into an unrestricted prompt dump.
| Capsule element | Runtime behavior |
|---|---|
| Resolved Business Unit | Use configured Business Unit/source-binding context first; ask only when client signals conflict or are materially ambiguous. |
| Compact facts | Include a small number of source-backed fact summaries with confidence and provenance. |
| Processing folder | Represent missing or stale folder context as a background system refresh, not a consultant management chore. |
| Communications context | Use compact communication summaries for objectives and caveats; do not treat comments as sourced facts. |
| Tool pointers | Expose deeper retrieval through Business Unit memory, source bindings, processing-folder context, communications context, report capsules, Asana thread summary, and fresh channel evidence. |
| Executable context read | The general-task workbench can run box.client_context_read for resolved Business Units with a current context pack or active Context Source Binding, returning compact source pointers and counts. |
| Maintenance logging | Memory gaps, processing-folder refresh needs, and communications-context refresh needs are logged as D1 run steps and deduped platform issues without adding consultant comment noise. |
| Asana visibility | The Delegation Confirmation can show capsule metadata, but not raw chunks, full reports, provider payloads, credentials, or long threads. |
The intake capsule remains feature-flagged with TEAM_AGENT_AGENCY_BRAIN_CAPSULE_ENABLED. The workbench context lane is now executable for compact D1/R2 retrieval; live Asana smoke tests still require explicit operator approval.
Conflicts
Numbers on their own don't tell you much. A rising CPA can be exactly what you signed up for when the client approved scaling spend; a conversion drop may mean nothing until lead quality is checked. Jessie calls out conflicts like these instead of smoothing them over, so you decide what they mean.
- If task notes conflict with approved context, Jessie should mention the conflict and ask for clarification when it affects the conclusion.
- If data conflicts with context, Jessie should explain the mismatch and suggest follow-up checks.
- If context is missing, Jessie should say what context was unavailable instead of pretending to know it. A Context Pack Version ID, manifest path, or generated Agent Context index alone is provenance, not reviewed context.
- If a consultant wants a task note to become durable context, it needs a separate context update path.
Governance
- Box Agent Context folders remain the consultant-maintained source of truth for client context documents.
- The Admin Registry Business Unit Context tab is the source of truth for which Box folder Agency OS should ingest.
- A context refresh is required before new or changed Box files become prompt-ready runtime context.
- Generated Agent Context indexes are metadata-only locator files; approved context belongs in real source documents.
- R2 and D1 provide prompt-time context retrieval, not an unrestricted model memory store.
- AI Gateway payload logging remains bounded by policy; D1 stores compact run-step metadata rather than raw provider payloads.
- Generated output can propose context updates, but Jessie must not silently rewrite trusted context.