The Agents list
- Stat cards: Total Agents, Active, Departments, Total Runs.
- Toolbar: search (“Search name, ID, dept, role…”), department filter, status filter (Active / Paused / Draft), Group by department, and a table/grid view toggle.
- Table columns: Name, ID, Department, Role, Model, Schedule (“Manual” when none), Status. Columns are sortable and drag-resizable (widths are saved per user).
- Row actions (on hover): ▶ Run now (active agents only) and a ⋯ menu with Edit, Duplicate, and Delete. Careful: Delete removes the agent immediately, without a confirmation dialog.
- Banners warn about failed runs in the last 24 hours and a low AI-provider balance.
Creating an agent
Three ways:The agent page
The agent page has the editor on the left and a Run Agent panel on the right. Header controls:- Execution type — Autonomous (the model decides on its own, using the briefing and tools; best for open-ended tasks) or Workflow (deterministic steps that run in order; best for structured pipelines).
- Quick model switch, Configure with AI (describe a change in plain English; you review a before/after diff and click Apply), and Save Changes.
Overview
Stat cards (Total Runs, Success Rate, Avg Response, Cost/30d) plus:- Agent Avatar & Icon — pick a face, icon, and color theme.
- Agent Details:
- Decision Bounds — the agent’s personal guardrails: “Ask me before the agent… spends more than $X” and “…does something irreversible (delete, mass-send, pay out)”. Both default to on. Otherwise the agent decides on its own. (Enforcement on autonomous tool calls is still rolling out.)
Briefing
The agent’s System Prompt — loaded on every run; it defines who the agent is, what it has, and how it behaves. Use the Auto-Generate / Improve button to draft or refine it with AI (review, then Append or Replace). For one-off tasks use the Per-Run Input in the run panel instead. In Workflow mode the system prompt is not used and the field is disabled.Skills & Tools
Everything the agent is allowed to use. Grants save automatically, and ungranted tools are blocked server-side — even if the model tries. Sections:The data ceiling
For each granted table you can open a per-source preferences editor (editable by admins; members see it read-only):- Columns this agent may read — unpicked columns cannot be requested by the agent at all, even if it asks.
- Default filters — always applied; the agent cannot opt out.
- Row limit cap — max rows per query (up to 50).
Workflow Steps (Workflow mode)
Instead of thinking autonomously, a Workflow agent runs a deterministic pipeline, top to bottom. You can drag to reorder, edit, or delete steps, and each step declares its input/output variables. 15 step types are available: Tool, Extract, Filter, Pick, Map, Split, Format, LLM, Approval, Loop, Skill, User Choice, User Form, Inbox, Knowledge.- The Extract step bridges raw tool output to the next step in plain English — no JSON wrangling needed; an inline “Insert Extract step” button appears where one would help.
- The Approval step pauses the workflow and sends a Telegram approval request; it continues only if approved.
- Generate with AI builds or modifies the whole workflow from a description; Test Run executes real tool reads (no delivery) and shows per-step input/output with timings.
- Version History tracks each saved workflow version with one-click Restore.
Triggers
Enable one or more triggers — the agent wakes on every active one:
Agents in the same reporting chain can also hand work to each other’s inbox mid-run — no separate trigger needed.
Output
Where results go after a run:- Target Document — the agent maintains a living knowledge document instead of a new one per run: a single document, or one per user / department / channel / custom key. Attach an existing document or create one on the spot.
- Delivery Channels — Inbox (team routing): results route automatically up the hierarchy (worker → supervisor → CEO → you). Telegram delivery is configured on the Telegram tab.
- Webhook Delivery — POST the run result to external URLs (HMAC-signed), per-event: run success, run failed, budget exceeded.
- API Actions — pre-configured external API calls (method, URL, vault credential fixed) that become callable tools for the agent; each has a Test button.
- During-Run Inbox — allow the agent to hand work to another agent’s inbox mid-run.
Telegram
Everything Telegram in one place: connect a bot (bot token lives in Settings › Vault), enter or Detect your Chat ID, and use the Send Test button to verify the wiring. Trigger modes — how incoming chat messages start the agent:
Voice messages — with accept voice enabled, voice notes are transcribed automatically and the transcript becomes the run input; an optional confirmation step asks before the run actually starts.
Talking back — enable result delivery to Telegram to have run results posted to the same chat, and toggle the during-run tools: Send Telegram Message (a non-blocking update mid-run), Ask User (blocking) — the agent posts a question with inline choice buttons and the run pauses until you tap one (or the prompt times out) — and Request Approval (non-blocking yes/no). Actions that require your sign-off arrive the same way: a message with inline approve / reject buttons, decided with one tap.
One voice. Only the top-level agent of a run talks to you on Telegram. When it delegates work to sub-agents, they report their results back up the chain, and you receive one coherent answer instead of several bots messaging you at once. Keep blocking prompts in mind for scheduled runs: a run waiting on Ask User stays paused until someone taps or the prompt times out.
Runs & Cost
- Savings — record how many human minutes and what hourly rate one run saves, to track ROI per run.
- Credit consumption — the agent’s billing policy and an optional max-credits-per-run override.
- Run budget — a hard USD ceiling per run; the runner stops the agent at the cap.
- Recent Runs — status, trigger, duration, cost, savings, tokens. Click a run to open the Run Audit Trail: trigger input, agent output, base system prompt, injected skills, registered tools, context sources, inbox tasks, user replies, conversation history, unverified claims, and the full assembled prompt — plus Replay to run it again. View All Runs jumps to the Runs view filtered to this agent; Reset History clears the stats.
History
Every briefing, prompt, or configuration change is captured as a version. Expand a row to see what changed, and Restore any previous version.The Run panel
On the right of every agent page: the agent’s status, an optional Per-Run Input (an ad-hoc prompt for this run only — it does not change the system prompt), the Run Now button (or Run Workflow (N steps)), recent runs, and quick stats. The agent must be Active to run.Hierarchy
Where: Sidebar › Agents › Hierarchy. A read-only org chart of your agents: stat cards (Total Agents, CEOs, Supervisors, Workers, Departments), three views (Department / Grid / List), and sorting by name, department, or role. Department panels group agents into Leadership / Supervisors / Workers bands; each agent card shows its live status dot, role badge, and success-rate bar, with run stats on hover. Click any card to open the agent. Reporting lines are edited on the agent’s Overview tab (Reports To Agent), not here. A warning appears if a circular reporting chain is detected.How delegation works
The hierarchy is not just a chart — it is the delegation protocol your agents actually follow:- Reports-to chains. Each agent’s Reports To Agent setting forms the chain: workers report to a supervisor, supervisors to the CEO agent, and the CEO reports to you (its digests land in your Inbox). After a run, an agent’s output travels one step up this chain automatically.
- Handing off work. An agent can put a task into another agent’s Agent Inbox — mid-run or as its result. The receiving agent does not wake up instantly: it picks up unread handoffs at the start of its next run (they are injected as “Pending tasks in your inbox — address these first”). For an immediate start, use the Handoff (on success) trigger instead.
- Delegated runs are children. When an agent spawns or delegates to another agent within a run, the delegated run appears as a child of the parent run in the Runs view — so one task fanned out across three agents still reads as one traceable tree, with each child’s cost and outcome attributed to the parent.
Permissions summary
Notes & not-yet-available
- Slack trigger requires additional Slack app setup and is not fully available yet.
- Decision Bounds enforcement on autonomous tool calls is still rolling out.
- Duplicate currently opens a blank create form rather than a pre-filled copy.
- There is no archive for agents — use Status: Paused; Delete is permanent and immediate.