Agents

Capabilities

The Agents page (/agents) is the builder desk. Create agents and automations from a description. Advanced on this page is detail, canvas, and executions — there is no separate Agent Studio window. Each agent has a Voice setting for phone calls and live voice; tools during a call still use the main model. The model you set is the one that runs: if its provider is not set up, the picker says so (the option reads not set up) and no other model runs in its place. Connect is account sign-in; Assign puts an app on this agent; Ready is both plus selected actions. Customer-facing phone agents live under Business.

• Three zones — orange My agents / My automations, center chat or recipe, green Connected services (External / Ecosystem / Channels)

• Same server as OS staff — they run on your Main Agent

• Advanced — agent detail (Prompts, Services, Triggers, Memory, Logs, Approvals), plus Steps / Canvas / Executions for automations

• Not Business phone agents — customer-facing voice and SMS live under Kontext Business

• Not app workers — those mint when you deploy an app, not with + New agent

What is the Agents page?

The Agents page (/agents) is the builder desk. Create agents and automations from a description, then turn them on. A blank canvas is the advanced path after something exists — not the only path. They run on your Main Agent — the same server as OS staff. There is no separate Agent Studio window. Advanced inside this page is full detail, canvas, and executions.

The three zones

Left (orange): My agents and My automations. Platform staff sit in a small presence strip, not mixed into My agents. MAIN is the runtime’s front door. Center: talk to the selected agent, or the automation recipe (the step list). Right (green): Connected services for that agent — External (Gmail and the rest), Ecosystem (built-in Kontext / KESA), Channels (this profile’s line). Hero Active means active agent profiles, not “automations are on.” Studio settings (Reach, Texting, Server) is the account number and host — not one agent’s card.

Connect vs assign vs Ready

Connect signs the account in (one Gmail for the whole Kontext account). Assign puts that login on this agent. Until you assign, the rail says On account — not assigned — that is idle, not broken. Ready means signed in, assigned, and at least one action selected. Remove from agent keeps the account login. Ecosystem tiles say Add to agent / On agent — no OAuth. If no module says On agent, that person cannot use any built-in Kontext area. Custom MCP is Advanced → Services, not an External tile.

Automations, canvas, and executions

An automation is a named list of steps (usually “this profile, with this input”). After you create it, Steps is the readable list, Canvas is the picture, Spec is read-only, Executions is run history. Run now starts one execution. Chat is a channel, not a trigger. Triggers you can create on a profile: scheduled, condition, webhook. Business recipes (/business/automations) are a different desk even when they share Main Agent.

Advanced, MCP, failed runs

Full agent: badge Advanced. Groups: Home, Configure (Prompts / Services / Triggers), Channels, Memory, Activity (Logs / Approvals), Settings, Lobby (Phone and inbox). Services → Advanced configuration → MCP configuration: endpoint, Claude API key, Composio key, MCP Tokens, Save MCP Configuration. Not Developer Gateway. Failed work: Prompts Failed, Runtime logs, Executions Runs → Watch a run → Pause / Resume / Stop / Rerun. Talk does not invent the reason — open the log.

What it is not

It is not Kontext Business. Customer-facing phone and SMS people live under /business/agents. It is not the Roster on My Kontext (people you hired). It is not an app worker — those mint when you deploy an app. Workspace → Agents is what chat sees on a project plus connect tools — not this studio. Agent Memory is this profile’s facts, not OS Memory Providers and not staff reports.

The business number is not your Twilio keys

You buy one number with Kontext credits. A second purchase is refused: You already have a business number. The same card is on Business lobby, OS Channels, Agents settings (Reach), and Profile → Reach — not three products and not Dana’s private line. A Twilio subaccount is a phone account Kontext creates for your A2P paperwork. Profile → API keys does not take a Twilio SID. Who answers is a claim, then the operator. US texts need A2P; voice does not. A2P “campaign” is carrier paperwork, not a Network challenge.

Kontext lets you create AI agents (standalone workers) and automations (multi-step workflows) on one Main Agent runtime. Two surfaces plus a project powers hub:

| Surface | Route / entry | Who it is for | What you do here | |---------|---------------|---------------|------------------| | Agents | /agents, module switcher, OS dock | Everyone — default home | Recipe-first studio: agents & automations, Connected services (Ecosystem then External), link to projects. Advanced on this page is detail, canvas, executions. | | Workspace → Agents tab | Inside a project | Project context | Powers hub — What chat sees + Connect tools; not the full builder |

There is no separate Agent Studio OS window. Power editing is Advanced inside /agents.

The Agents module (/agents) creates agents and automations on your unified runtime. Recipe-first builder, Connected services (Ecosystem then Composio), and link to project for Workspace chat awareness. For in-platform actions see KESA. Advanced inside /agents is full detail, canvas, and executions — there is no separate Agent Studio OS window.

• Left rail — My agents, my automations, + New agent, + New automation

• Unified runtime — Profiles on Main Agent (~540 hosting credits once for a new pair)

• Entity orchestration — Same pipeline from /agents and OS Start Building

• Not Business phone agents — CRM voice/SMS agents live under Kontext Business

Kontext lets you create AI agents (standalone workers) and automations (multi-step workflows) on one Main Agent runtime:

Platform agents are not Kontext Business phone/SMS agents (/business/agents).

Routes: /agents · Workspace → Agents (powers hub)

What the agent platform is

Kontext lets you create AI agents (standalone workers) and automations (multi-step workflows) on one Main Agent runtime:

Platform agents are not Kontext Business phone/SMS agents (/business/agents).

Agents module layout (Zone Split studio)

Unified creation (entity orchestration)

All surfaces share one pipeline:

Describe → discover runtimes (Main Agent first) → generate spec → connect KESA + Composio → deploy → workspace registry → optionally link to project (chat awareness, app integration).

Surfaces: Agents module, OS Start Building when it routes here. Canvas and executions are Advanced inside /agents.

KESA & ecosystem routing

Agent types & deployment

MCP & integrations

Triggers, monitoring, safety

Scheduled, webhook, conditional, manual, event-based triggers. Metrics, logs, approvals, retry/fallback, fuel preflight before deploy.

Agent ↔ project integration

Link to project wires chat awareness (aiAwareAgentIds), runtime profile links, optional in-app chat components, workflow hooks. Unlink removes project access; workspace agent remains.

OpenRouter BYOK

Profile → API keys: optional OpenRouter override routes workspace AI through your key and chosen model (workspace picker shows routing chip when active).

Advanced inside /agents

Agent groups: Lobby (Main only), Home, Configure (Prompts, Services, Triggers), Channels, Memory, Activity (Logs, Approvals), Settings. Automation views: Steps, Canvas, Spec, Workflows, Executions. Studio settings: Reach, Texting, Server.

Module relationships

User journeys

AM-01. Open the Agents module

Path: Module switcher → Agents (or /agents)

You do:

1. Sign in to Kontext.
2. Open the module switcher in the header or OS dock.
3. Choose Agents — URL lands on /agents.
4. Scan the Zone Split studio: orange left rail (My agents / My automations), center composer, green right rail (Connected services).
5. Note the health footer on the left rail for runtime status.
6. Confirm this is the recipe-first home — not Business phone agents. Canvas and executions are Advanced on this same page, not a second window.

Result: The Agents module opens ready for plain-language create flows with KESA + Composio on the right rail.

AM-02. Create an agent with a plain-language recipe

Path: /agents → + New agent → center composer

You do:

1. Click + New agent in the orange left rail.
2. Type what the agent should do in plain language in the center recipe composer.
3. Submit — entity orchestration generates a spec (name, instructions, suggested services).
4. Review the generated summary and Connected services suggestions on the green right rail.
5. Enable Kontext Ecosystem (KESA) and toggle module scope chips; connect Composio apps after KESA.
6. Confirm deploy to Main Agent (credit consent only if no initialized runtime exists).
7. Save — the agent registers in your workspace registry and appears under My agents.

Result: A new agent profile on the unified runtime, listed in /agents and visible to OS delegation when global or linked.

AM-03. Review the generated spec before you deploy

Path: /agents → + New agent → review step

You do:

1. Click + New agent and describe the job in plain language.
2. Pick Main Agent (or consent to a new server if none exists).
3. Wait for the generated spec — name, instructions, suggested services.
4. On Connected services, turn on Kontext Ecosystem modules this agent should use inside Kontext.
5. Connect or assign any required External apps until they say Ready (or skip and do it after deploy).
6. Confirm deploy.
7. Find the new row in My agents.

Result: You saw the spec before a profile existed — not a silent create, and not a template picker as the only path.

AM-04. Connect KESA on a new profile

Path: /agents right rail → Connected services → Kontext Ecosystem

You do:

1. During or after agent create, open Connected services on the green right rail.
2. Open the Ecosystem mode — built-in Kontext modules, no OAuth.
3. Confirm Ecosystem is listed as its own rail (not mixed into Gmail tiles).
4. Tap Add to agent on the modules this person should search (Marketplace, Finance, Workspace, Business, …).
5. Open Advanced → Services if you need a tighter allowlist.
6. Send a test message from the center chat that needs an in-platform action (list projects, a lead, deploy status).

Result: The agent searches and executes Kontext-native actions before any external OAuth tool is considered.

AM-05. Connect Composio during agent create

Path: Connected services → external app → Connect

You do:

1. With KESA enabled, scroll Connected services below Kontext Ecosystem.
2. Pick an external app (Gmail, Slack, Stripe, Notion, …) the spec requires.
3. Click Connect if the account is not signed in, then Assign if the rail says On account — not assigned.
4. Optionally restrict the Tools allowlist (Advanced → Services) until the tile says Ready.
5. Repeat for additional apps the recipe flagged as required.
6. Finish deploy.
7. Verify tiles say Ready (green), not only “connected.”

Result: External SaaS tools are available to the agent after KESA — use when Kontext cannot perform the action natively.

AM-06. Scope KESA modules per agent

Path: Agent card → Connected services → module chips

You do:

1. Select an existing agent in My agents.
2. Open Connected services → Ecosystem.
3. Ensure modules you want show On agent; tap Add to agent on the rest this person should use.
4. Leave Finance (or any other module) off if this helper should not touch it.
5. Open Advanced → Services if you need a tighter list.
6. Run center chat with a task that would require a disabled module — it should not pretend it ran.
7. Add the module later if the job truly needs it.

Result: Per-profile least-privilege KESA scope — same chip model as OS Employees and automations.

AM-07. Link an agent to a project

Path: Agent card → Link to project

You do:

1. Open the agent you deployed from /agents (workspace registry entry).
2. Click Link to project and pick the target project from the list.
3. Confirm chat awareness options — agent joins aiAwareAgentIds for Workspace Chat when enabled.
4. Review optional in-app integration hooks shown in the link dialog.
5. Save the link.
6. Open that project in Workspace → Agents → What chat sees and verify the agent can be turned on for Chat.

Result: Workspace Chat and project-scoped hooks can invoke the agent; the workspace registry entry remains the source of truth.

AM-08. Keep an agent available beyond one project

Path: /agents → create or leave unlinked · Workspace → What chat sees

You do:

1. Create an agent from /agents without linking a project, or unlink later (AM-20).
2. Confirm it still appears under My agents.
3. Open a project → Workspace → Agents → What chat sees and turn the agent on for that project's Chat when you want it there.
4. Repeat for a second project if Chat should see the same person.
5. Do not expect + New agent to mint an app worker — that happens at app deploy (AM-26).
6. OS staff stay in the presence strip; they are not this list.

Result: Studio agents live on the account. Project Chat only sees who you turned on under What chat sees.

AM-09. Create an automation from /agents

Path: /agents → + New automation

You do:

1. Click + New automation in the orange left rail.
2. Describe the workflow in plain language (who does what, in what order).
3. Review the generated step list — each step should name a profile.
4. Connect Ecosystem / External on any participating agent that still shows not Ready.
5. Confirm create on Main Agent.
6. Find the workflow under My automations.
7. Open the center recipe and Run now once, or open Advanced Canvas / Steps to edit.

Result: Automation is in the account registry and listed in /agents. Canvas is Advanced after create, not a create-mode picker.

AM-10. Run an automation from the center recipe

Path: /agents → select an automation → center recipe → Run now

You do:

1. Select an automation in My automations.
2. Read the step list in the center recipe — names of agents, order of work.
3. Turn the automation on if it is off.
4. Use Run now (add optional run input if the recipe asks).
5. Watch live progress in the center, or open Executions.
6. If a step fails, open the named agent’s Connected services (Ready vs On account — not assigned).
7. Re-run after the fix.

Result: One execution without opening Canvas. Same workflow as Steps / Canvas.

AM-11. Open Canvas after an automation exists

Path: /agents → automation → Canvas (Advanced)

You do:

1. Select an automation and open Advanced (or the recipe’s open-canvas control).
2. Use the Canvas tab — boxes and lines for the same steps as Steps.
3. Add, remove, or reconnect steps; sequential / parallel / conditional as supported.
4. Open a step to confirm it names a profile that is Ready for the tools it needs.
5. Switch to Spec only to read the outline — you do not edit there.
6. Run now or open Executions for the first run.
7. Return to the studio list via Workflows.

Result: Graph edits persist on the same Main Agent workflow. You did not start from a blank canvas at + New automation.

AM-12. Open Advanced agent detail

Path: /agents → agent → Advanced (or ?view=agent-detail)

You do:

1. Select an agent in My agents.
2. Open Advanced — groups: Home, Configure, Channels, Memory, Activity, Settings. Lobby only on MAIN.
3. Configure → Prompts to run a task now; Services for KESA, Composio, custom MCP; Triggers for schedule / condition / webhook.
4. Channels for this profile’s line (claim, email mint) — not the company buy-number card.
5. Memory for this profile’s facts (not OS Memory Providers).
6. Activity → Logs and Approvals. Settings for model, require-approval, remove.

Result: Full profile desk inside /agents. No second OS window.

AM-13. Edit a workflow on Canvas

Path: /agents → automation → Advanced → Canvas

You do:

1. Open the automation’s Canvas tab inside /agents.
2. Add, remove, or reconnect steps; update sequential / parallel / conditional edges.
3. Open participating agents to adjust Ecosystem chips or External assign/Ready if steps fail.
4. Save canvas changes.
5. Run from the canvas or Executions.
6. Confirm behavior in Logs / the execution timeline before enabling a schedule or webhook.

Result: Graph-level edits persist without recreating the automation.

AM-14. Watch executions

Path: /agents → Advanced → Executions (?view=workflow-executions)

You do:

1. Open Executions from the automation Advanced tabs.
2. Open a recent run — step timeline, what ran, what failed.
3. Retry or approve when the UI shows a gate (Approvals on the agent if it is waiting on you).
4. Cross-check failures against the agent’s rail: not Ready, Assigned — no tools selected, Needs re-auth, Ecosystem module off.
5. Fix, then re-run from Executions.

Result: Run history lives on /agents, not a separate studio window.

AM-15. Configure OpenRouter BYOK

Path: Profile / workspace settings → API keys → OpenRouter

You do:

1. Open your profile or workspace API keys settings (not Business phone agent settings).
2. Locate OpenRouter optional override for workspace AI routing.
3. Paste your OpenRouter API key and choose the model you want routed through BYOK.
4. Save — the workspace model picker shows a routing chip when BYOK is active.
5. Return to /agents and create or test an agent using workspace AI paths that honor BYOK.
6. Confirm billing runs through your OpenRouter account for those routed calls.

Result: Workspace AI (including agent test chat where wired) can use your key and model instead of platform defaults only.

AM-16. Test an agent from the module chat panel

Path: /agents → agent card → Test or inline chat panel

You do:

1. Select the agent in My agents.
2. Open the Test control or inline chat panel on the center/right zone.
3. Send a realistic task that exercises KESA (in-platform action) first.
4. Send a follow-up that requires Composio if connected — watch the trace for tool routing order.
5. Verify execute with routing enriches the task with ecosystem context before execution.
6. Fix Connected services if the tile is not Ready (assign, allowlist, or re-auth); re-test.
7. Link to project (journey AM-07) only after the test is useful.

Result: Confirms KESA scope, Composio OAuth, and model routing before production use in Workspace Chat.

AM-17. Follow unified entity orchestration deploy flow

Path: Any create entry (/agents, OS native create) → describe → deploy

You do:

1. Start create from /agents + New agent or + New automation (or OS Start Building when routed here).
2. Describe the entity in plain language — orchestration discovers runtimes (Main Agent first).
3. Wait for generate spec — review name, instructions, workflow steps, dependency hints.
4. Connect KESA + Composio on the right rail; complete OAuth in the dependency wizard if prompted.
5. Accept deploy to Main Agent — credit consent only when no initialized runtime exists (~540 for new pair).
6. Confirm workspace registry write — agent or workflow appears in account-level registry.
7. Optionally link to project for chat awareness and app integration hooks.

Result: Same pipeline from /agents and OS Start Building — one runtime, one registry.

AM-18. Toggle what Workspace Chat sees

Path: Workspace → Agents → What chat sees

You do:

1. Open a project in Workspace.
2. Go to Agents — tabs: What chat sees and Connect tools.
3. Turn on agents and automations Chat may use on this project.
4. Connect tools stays disabled until at least one agent is on under What chat sees.
5. Do not expect full studio create here — use /agents for + New agent and Canvas.
6. Open Connect tools only after Chat can see someone — that attaches Composio to chat-aware agents.
7. Return to Workspace Chat and mention the agent by name.

Result: Project Chat respects What chat sees. Create, canvas, and executions stay in /agents.

AM-19. Connect tools from Workspace Agents

Path: Workspace → Agents → Connect tools

You do:

1. Open Workspace → Agents for the project.
2. Confirm at least one agent is on under What chat sees.
3. Open Connect tools — this is the Workspace Integrations agent-only attach, not /agents Advanced Services.
4. Connect or attach apps those chat-aware agents should use.
5. If a tile is only on the account, assign it the same way as the studio rail (until Ready).
6. Run a Workspace Chat task that needs the newly connected service.

Result: Project-context attach without duplicating the agent. /agents remains the builder desk.

AM-20. Unlink an agent from a project

Path: Agent card → Link to project → Unlink (or Workspace Agents → unlink)

You do:

1. Open the agent in /agents or Workspace → Agents.
2. Open Link to project management for the current link.
3. Choose Unlink from the target project.
4. Confirm — aiAwareAgentIds and project-scoped hooks remove access for that project.
5. Verify the agent still appears in My agents at workspace level.
6. Workspace-global agents (All projects) remain available elsewhere; project-only agents simply lose that project's Chat hooks.

Result: Project loses Chat and integration access; the workspace agent profile and registry entry remain intact.

AM-21. Open the automations dashboard

Path: /agents → Advanced → Workflows (?view=workflow-dashboard)

You do:

1. Open Advanced workflow views from an automation or the Workflows tab.
2. Scan counts and cards — this is the list, not a retired Digest window.
3. Jump into Canvas, Steps, or Executions from a card.
4. Compare with My automations in the orange rail.
5. Hide platform staff workflows — they are not in this list unless you deep-link one.
6. Check the left-rail health footer if the runtime looks down.

Result: Account workflow list inside /agents.

AM-22. Retry a failed execution from logs

Path: /agents → Advanced → Logs or Executions → failed run

You do:

1. Open agent Logs or automation Executions.
2. Select a failed run — read the error (Ecosystem denied, expired login, fuel, waiting on approval).
3. Fix on the profile: Assign / Ready, widen Ecosystem, Approvals, or add fuel.
4. Retry or Run now again.
5. Confirm the retry completed.
6. Turn a schedule or webhook back on only after a clean run.

Result: Failed automations recover without rebuilding from + New automation.

AM-23. Set triggers on an agent

Path: /agents → Advanced → Configure → Triggers

You do:

1. Open the agent’s Triggers pane.
2. Create scheduled, condition, or webhook — those are the create buttons. Chat is a channel, not a trigger.
3. Workflow schedules (on the automation) are a different desk from per-agent triggers.
4. Set retry or limits when shown.
5. Save. Confirm the runtime is healthy.
6. Wait for or fire the trigger; the run appears under Executions.

Result: Profile triggers fire that agent. Not Business SMS/voice recipes.

AM-24. Create from OS Start Building (routed to entity orchestration)

Path: Kontext OS → Start Building → Agent or Workflow

You do:

1. Open Kontext OS → Start Building (banner, Start menu, or ⌘⇧B).
2. Click Agent or Workflow — describe in plain language.
3. Follow spec generation, review, service authorization, deploy — Main Agent first.
4. Ecosystem before External, same as /agents.
5. On success, open /agents — the entity appears in My agents or My automations.
6. Open Advanced if you need Canvas or the full Services desk.

Result: OS-native entry into the same create pipeline as the Agents page.

AM-25. Distinguish platform agents from Business phone agents

Path: /agents vs /business/agents (awareness only)

You do:

1. Open /agents — builder desk for studio agents and automations on Main Agent.
2. Note Connected services: Ecosystem + External for operator/builder tasks.
3. Open /business/agents separately — customer-facing phone/SMS/voice people.
4. Confirm Business agents use Business integrations — not the /agents left rail.
5. Create builder agents from /agents or Start Building — never from Business phone wizards for code work.
6. OS platform staff and Business digital employees are two more lists (not this studio).

Result: /agents is the builder. Business is the company. Same Main Agent family, different desks.

AM-26. Tell an app worker from an agent you linked

Path: Hub → Workforce · Workspace → Agents · /agents

You do:

1. Deploy an app (or open one already live).
2. Open Hub → Workforce — the dedicated app worker for that app is a persistent profile minted at deploy (kontext-app.json).
3. Open Workspace → Agents on that project — the dedicated worker is not the same row as an agent you linked yourself.
4. Linked agents live as attachments you chose. The worker stays until the app is removed.
5. Bind email or a phone number on the worker's Channels card if the app should answer mail or calls.
6. Do not fire platform staff looking for this worker — staff live on Employees → Platform.

Result: You can tell deploy-minted workers from studio-linked agents and from OS teammates.

AM-27. Buy the account number, register A2P, then bind who answers

Path: Business lobby · /agents settings (Reach / Texting) · OS Channels · not nested inside one agent row

You do:

1. Open Business → lobby, OS → Channels, or /agents settings (Reach) — the buy-number card is the same account card. It is not inside Dana.
2. Search an area code, buy, then Verify wiring. Rental bills Kontext credits, not Gas Station.
3. Register A2P 10DLC (brand, then campaign). On Business the form is on the lobby card. In /agents open Texting. Calls work without A2P; US texts do not.
4. Wait if the chip says With the carriers (often 1–7 days). That is review, not a broken agent.
5. On the Channels rail or Advanced → Channels, claim the number on the profile that should answer. Operator/reception is a separate designation.
6. Do not use /business/agents as a substitute for the builder studio, or the studio as a substitute for the Business lobby.

Result: One account number, one A2P record, and an explicit answerer. SMS that “does nothing” while calls work is A2P or wiring first, not a missing agent.

AM-28. Connect vs assign vs Ready on the green rail

Path: /agents → select agent → Connected services → External

You do:

1. Select Dana (or any studio agent).
2. Open External. If Gmail is missing, Connect (account OAuth — one login for the whole account).
3. If the tile says On account — not assigned, Assign — that is idle, not an alarm.
4. If it says Assigned — no tools selected, open Advanced → Services and pick actions.
5. Stop when the tile says Ready.
6. Remove from agent later if you want — the account login stays. Disconnect on Connected apps if you mean revoke the account.

Result: You can explain the three jobs. Connected is not Ready. Ecosystem tiles never OAuth.

AM-29. Open Agent Memory without mixing it up with AME

Path: /agents → Advanced → Memory

You do:

1. Select a studio agent (not a read-only system profile if you need to add facts).
2. Open Memory — facts, pins, recent events for this profile.
3. Add a fact if you want this person to remember it.
4. Do not look here for OS staff reports, Today brief cards, or Memory Providers (Mem0/Notion).
5. Remember AME still routes where a filed memory is stored on the account.

Result: Profile memory is a list on this agent. The OS brain and AME are different desks.

AM-30. Choose the voice model for calls

Path: /agents → an agent → Voice

You do:

1. Open the agent.
2. Open Voice (the control is labeled Voice model).
3. Pick a live-voice model. Phone calls and live voice use this model.
4. Tools during a call still use the agent’s main model.
5. Clear the choice to return to the platform default.

Result: Calls and live voice use the model you picked. Text work stays on the main model.

How it works

Route /agents. Agent Studio / recipe-first automations; shared entity orchestration with OS create-agent. MCP/Composio; KESA routing; runtimes + Zoe where needed.

Benefits

• Always-on workers without a separate automation SaaS.
• Same fabric powers OS employees and Business digital employees.

Who it's for

Primary: builders and operators automating work.