Module
Agentic AI
Agents that do work, the workflows that orchestrate them, and the systems they reach. An agent is not a chatbot and not a batch job: it chooses its next step at run time, against tools it was granted, under limits somebody set. Every one of those choices is recorded — which model, which version of which prompt, which tools it held, what it called, what came back, what it cost and who approved it — because "the agent did it" is not an answer anybody can audit.
6 groups · 9 sections · 29 screens
Build & Orchestrate
What an agent is, what it may call, and the canvas it is joined up on.
RecordsAgentsWhat each agent is for, what it may do, and what it has actually been doing. An agent with no tools is not broken — it is one that can only think, which is a legitimate and common thing to want.OpenRecordsTool GrantsWhich agent may call what, and on what terms. The same tool is a different risk in two agents' hands, which is why the limits live on the grant rather than on the tool.OpenRecordsKnowledgeWhat an agent may read that is not a tool call — a folder, a set of articles, a report. Retrieval rather than action.OpenViewAgent VersionsEvery published prompt, with what it was allowed to call. A run names the version it used, so a decision made three months ago reads against the instructions that were live then rather than the ones live now.Open
WorkspaceAgent StudioThe canvas. Drag agents, tools, decisions and human gates onto it, join them, and publish when it validates.OpenRecordsAgent WorkflowsThe graphs. What runs is the PUBLISHED SNAPSHOT, not what is on the canvas — so editing a workflow while a run is in flight changes nothing about that run.OpenViewPublished VersionsWhat each published version of a graph actually was. A run in flight keeps running against the version it started on.OpenRecordsTriggersWhat starts a run. A trigger created from a template is never armed — a new tenant should not have agents running before anybody has read what they do.Open
Reach
The systems outside this platform that agents can talk to.
RecordsAgent ConnectionsThe MCP servers, APIs, webhooks and other agentic platforms this company reaches. An MCP server describes its own tools, so the catalog is the server answer rather than somebody typing.OpenRecordsDiscovered ToolsWhat each connection turned out to be able to do. A tool the server no longer offers is disabled rather than deleted, because deleting it would take the grants and the call history with it.OpenRecordsTool CatalogEverything an agent may be granted, of every origin, in one list — because the model is given one list and the person granting authority should be reading the same one. A resource tool is the platform own endpoint, so it inherits the module grant, the org scope, the numbering, the editability rule and the audit trail unchanged.Open
Authority
The limits, and what broke them.
RecordsGuardrailsThe limits that are not the agent's own. Checked before every tool call and again against what it would move. A new deployment with agents and no limits is one bad prompt away from a thousand purchase orders.OpenRecordsIncidentsA guardrail broken, a loop caught, a cost overrun, an unsupported claim. Separate from the run because a breach outlives it: it is the thing somebody reviews on Monday.Open
Running
What is happening now, what is waiting on somebody, and what came out.
WorkspaceMonitorWhat the agents are doing, what has stopped for a decision, and the trace of any run read as the narrative it is rather than as four separate tables.OpenViewAgent RunsEvery run, with what it did, what it cost and what stopped it. A run is not a record somebody edits — it is the account of what happened.OpenViewAgent ApprovalsWhat is waiting on a person. A run stops here and survives a restart while it waits; an approval nobody answers expires rather than hanging.OpenViewDelegated TasksWork an agent handed to a licensed Claude worker because it was too big for one model turn. Claim one to get a worker token and the connection for Claude Code, the Agent SDK or a managed agent; the worker reaches this company only through the agent's own tool grants, guardrails and approvals, and every call is recorded.OpenViewWhat Agents ProducedEvery record created, message sent, proposal made and finding recorded. Accepting a proposal creates the record THROUGH THE API with your name on it, not the agent — which is where most of the value of a suggest-level agent actually lands.OpenViewTool CallsEvery call an agent made and what let it through. This column — grant, approval or guardrail — is what turns this table from a log into evidence.Open
RecordsEvaluationsWhether the agent was any good — graded by a person, by another agent, or by what happened to what it produced. A proposal accepted and still standing thirty days later is the strongest signal this table can carry.OpenRecordsAgent MemoryWhat each agent has kept between runs, and when it will forget it. Only where the agent memory mode allows it.OpenWorkspaceSentinelsThe plain checks that warn before a save -- a duplicate requisition, an invoice entered twice, a customer over their limit -- how often each warning was heeded, and the company's switch and thresholds for each.OpenReportCompany MemoryWhat the company has learned: warnings explained, approval and rejection reasons, outcomes worked out from the records, changes confirmed in ASK. Each row is shown only to people who may open the record it came from.Open
Reporting
Cost, value, and whether any of it is trusted yet.
ReportReports & AnalyticsThe report catalog for this module, and the models behind it — dimensions and key figures you can rearrange, filter, total and export.OpenReportPredictive & PrescriptiveWhere this module’s measures are heading, how much that projection is worth, and what to do about it while there is still time.Open
More
Reachable by link
3These screens carry no tile on the module page. They are reached from the record they belong to, from a saved view, or by their address — and they behave like any other list once open.
RecordsNodesEvery node of every graph, as a list. The canvas is the way to work on these; this is here for searching across workflows and for a saved view.OpenRecordsEdgesHow the nodes are joined and what decides which way a run goes.OpenViewStepsEvery node execution across every run. The trace on a run is the way to read these; this is for searching across runs.Open