Under the hood
One coach. A whole system behind it.
A coaching message in Slack looks simple. Behind it runs an agent built for one customer, a model of how that team sells, a sandbox for everything it executes, and a pipeline that turns every call into context. Twelve subsystems, each a product of its own. Follow one call through all of them.
01 / 12Connected tools
Your tools stay where they are
The CRM stays the system of record and the recorder stays the recorder. Taskbase mirrors deals, companies, contacts, calendars and documents into one model and never becomes a second place to maintain them.
- HubSpot or Salesforce stay the system of record; Taskbase mirrors what the coach needs and can write back
- Gong, Fireflies, Fathom, Zoom or Google Meet push a recording the moment a call ends; Taskbase’s own recorder joins calls only where a team wants it
- Calendars and documents too: Google Calendar, Microsoft 365, Notion, Confluence
- Custom integrations take days, not months
The call 14:32 · A discovery call with Helio Logistics ends on Zoom. The recording is pushed to Taskbase.
Next: The data layer02 / 12The data layer
One governed data set per customer
Everything the coach knows lives in a data layer of its own, one per customer: typed collections of records, files and documents, with access rules modelled on your organisation and an export at any time.
- Typed collections: conversations, deals, accounts, people, documents
- Access rules follow your org chart: a rep sees their own calls, a manager their team’s
- Reached as an MCP server by Taskbase agents, Claude, ChatGPT and your own apps
- An asset in itself: the whole revenue organisation in one governed data set your team can read and explore, and the base every app on demand is built on
The call The transcript lands in your organisation’s data layer, pseudonymised, linked to the deal, the account and the two people on the call.
Next: The revenue brainbrightline · data layer Export at any time
03 / 12The revenue brain
One model of how the team sells
The Revenue Playbook is the hub: how the team sells, what good looks like per role, the answers to objections. Around it sits one connected model of people, deals, conversations and skills.
- Taskbase’s own sales knowledge: MEDDICC, SPICED, SPIN, Gap Selling and other methodologies, expert playbooks and templates. The coach builds and fills your playbook from this library and your own material, so no team starts from a blank page
- The playbook changes only through suggestions a sales leader accepts
- Every conversation is linked to the deal and the people on it; transcripts can be pseudonymised before they are stored
The call The call is now part of the model: the Helio renewal, the two contacts, and the playbook for account executives.
Next: Routines and events04 / 12Routines and events
Work that starts without a prompt
Scheduled jobs and incoming events wake the coach: a briefing before every meeting, feedback after every call, a nudge when a deal goes quiet. Nobody has to ask.
- Routines run on a schedule, per team and per person
- Events from the connected tools start work the moment something happens
- Every run is a normal coach run, with the same permissions and the same memory
The call No one asked. A routine notices the new coached call and starts a run for Anya’s coach.
Next: The orchestrating agentRoutines · Anya’s coachmeetingsroutinesweek of 12 Oct
Mon 12
Tue 13
Wed 14
Thu 15
Fri 16
Not on the calendar: routines that wait for an event
on call.ended Call feedbackrunning
deal quiet for 10 days Deal nudgewaiting
A run like any other: the same permissions, the same memory. Nobody had to ask.
05 / 12The orchestrating agent
A run is planned, split and delegated
A run is a loop of reasoning and tool use. The orchestrator plans it, decides which tools and skills it needs, and hands long tasks to specialised sub-agents that work side by side and report back.
- Sub-agents are scheduled, queued and run in parallel, each with a task of its own
- Specialists for scoring a call, analysing a deal or drafting a follow-up
- The orchestrator collects the results and decides what the rep should hear
The call Three sub-agents start: one scores the call, one checks the deal, one drafts the follow-up.
Next: Skills, plugins and memoryOrchestrating agentrun 8f3a
- Check the call against the playbook for account executives
- Look at the Helio renewal: stage, history, risks
- Draft the follow-up Anya promised on the call
Call scorerscores the call against the playbook
queuedrunningdone · 12 s
Deal analystreads the Helio renewal’s history
queuedrunningdone · 9 s
Follow-up writerdrafts what Anya promised
queuedrunningdone · 7 s
Three results back. The orchestrator decides what Anya should hear.
06 / 12Skills, plugins and memory
Assembled, not prompted
Capabilities are packaged as versioned skills and plugins. A customer’s coach is assembled from them and from its own configuration, with a structured long-term memory kept apart from conversation history.
- Versioned skills and plugins, updated without touching a customer’s configuration
- Memory is structured and per customer: vocabulary, preferences, what worked
- Remembering a preference never means re-reading every chat
The call The coach loads the call-scoring skill and remembers that Anya wants one concrete thing to improve, not five.
Next: The sandboxAnya’s coachBrightline
- call-scoringv12skill
- playbook-aev7Brightline’s own
- crm-hubspotv4plugin
- meeting-briefv9skill
Assembled from versioned parts and its own configuration. Not one ever-growing prompt.
Long-term memory
Anya wants one thing to improve per call, not five.
“Discovery” means the first call at Brightline, not the demo.
A pricing anchor early in the call worked with logistics prospects last quarter.
Structured, per customer, apart from conversation history.
07 / 12The sandbox
Code runs isolated, then disappears
Anything the coach executes, from working through a transcript to building a report, runs in a sandbox created for that coach and session: isolated, with hard limits on memory and time, cleaned up afterwards.
- Arbitrary code execution that cannot reach the platform: no network except a proxy
- Data only through that proxy, with a short-lived, run-scoped token
- Hard memory and time limits; the sandbox is destroyed when the run ends
The call The call scorer writes code to work through the transcript. It runs in a sandbox that exists for twelve seconds.
Next: The model gateway$ score-call --run 8f3a --role account-executivereading transcript (pseudonymised) · 48 minloading playbook/account-executive · 6 skillsscoring discovery-questions … 71 (+9)scoring objection-handling … 50 (+2)writing 14 evidence records → data layer, via proxydone in 12.4 s
- Isolated
- No network, except a proxy
- Run-scoped token
- Hard limits on memory and time
- Disposable
Sandbox destroyed. Nothing left behind, not even the code.
08 / 12The model gateway
Model orchestration, in the EU
A gateway in front of several model providers orchestrates them: it picks the model for each task and fails over between providers, so the coach gets better as models do, without your team changing anything.
- Orchestration per task: long context, speed or reasoning decide the model
- Automatic failover when a provider is slow or down
- Inference in EU regions with zero data retention at the provider
The call The scoring task needs long context. The gateway routes it to one EU model and fails over when that one times out.
Next: Skills and scores09 / 12Skills and scores
Evidence, not opinions
Skills are a framework of observable sales behaviours. Every coached call adds evidence against it, and scores build up across calls rather than swinging on a single conversation.
- Scored against the skills the playbook defines for this role
- Every score comes with the moment in the call it is based on
- Rings move slowly on purpose: one call is evidence, not a verdict
The call Discovery questions move from 62 to 71. Objection handling barely moves: the pricing objection was answered, not explored.
Next: Interfaces and apps on demandDiscovery questions62 → 71
4 open questions in the first 10 minutes
Objection handling48 → 50
the pricing objection was answered, not explored
Next steps80 → 84
a dated follow-up agreed before the end
One call is evidence, not a verdict. Scores build up across calls.
10 / 12Interfaces and apps on demand
A screen when a chat is not enough
When a message is not enough, the coach renders a screen: a skill dashboard, a meeting briefing, a deal review, a form that updates the CRM. Whole apps too, assembled from the connected tools and the revenue brain.
- Built from templates your team can change
- Served with a strict content security policy: they show your data and cannot send it anywhere else
- A RevOps command center, a deal risk board or a cold-call list is assembled, not bought as a module
The call Anya’s skill dashboard is built from the new evidence. Her manager’s team view updates from the same data.
Next: Where people workAnya — Skill dashboard
Updated just now
Focus skillsChange focus
Top strengths
Problem impact & business case progress
3 Aug17 Aug31 Aug14 Sept28 SeptToday
Meetings per week
CW 32CW 33CW 34CW 35CW 36CW 37CW 38CW 39CW 40CW 41
▸ How the score works
11 / 12Where people work
One coach on every surface
One session model behind every channel. An adapter maps a Slack thread, a Teams chat, a web conversation or a phone call onto the same kind of session, so context follows the person, not the tool.
- Slack, Microsoft Teams, Google Chat, the web app and the phone
- Identity is resolved at the edge, before anything reaches a model
- Nothing is posted to channels: coaching is a direct message
The call 14:41 · One strength, one thing to improve, as a direct message in Slack. Nine minutes after the call ended.
Next: Identity, isolation and hosting
Taskbase CoachAPP14:41
Good discovery call with Helio, Anya. Strength: four open questions in the first ten minutes, and you let them talk.
One thing to improve: when they raised price, you answered it. Next time, explore it first: “What would make the price feel right?”
Open the momentPractice it
Slack%22%3E%3C%2Fpath%3E%0A%09%09%09%09%3Cpolygon%20fill%3D%22%23FFFFFF%22%20points%3D%2294.2080524%2095.1248514%2072.3884185%2095.1248514%2072.3884185%20154.540713%2058.4871168%20154.540713%2058.4871168%2095.1248514%2036.768788%2095.1248514%2036.768788%2083.5989345%2094.2080524%2083.5989345%22%3E%3C%2Fpolygon%3E%0A%20%20%20%20%3C%2Fg%3E%0A%3C%2Fsvg%3E%0A)
Microsoft Teams
Google Chat- Web app
- Phone
The same coach, the same memory, on every surface. Always a direct message, never a channel post.
12 / 12Identity, isolation and hosting
What a demo does not show
A prototype answers a question. A coach a whole revenue team relies on needs isolation between customers and between people, an identity on every request, and data that stays where it should.
- Single sign-on for people, short-lived identities between services
- Each customer has their own agent, configuration and data layer; each person sees only what they may
- Hosted in the EU and Switzerland, every processing location listed in the Trust Center
The call The whole run acted as Anya, inside Brightline’s organisation. It could read what she can open, and nothing from another customer.
Next: the whole systemThe whole system
Twelve subsystems, one coach
What worked on the call flows back into the playbook as a suggestion a sales leader can accept. The next call starts better. Every part above is a real part of the platform, built to run for a whole revenue team, every day.
- 01Connected toolsYour tools stay where they are
- 02The data layerOne governed data set per customer
- 03The revenue brainOne model of how the team sells
- 04Routines and eventsWork that starts without a prompt
- 05The orchestrating agentA run is planned, split and delegated
- 06Skills, plugins and memoryAssembled, not prompted
- 07The sandboxCode runs isolated, then disappears
- 08The model gatewayModel orchestration, in the EU
- 09Skills and scoresEvidence, not opinions
- 10Interfaces and apps on demandA screen when a chat is not enough
- 11Where people workOne coach on every surface
- 12Identity, isolation and hostingWhat a demo does not show
See it on your own calls
Upload three recorded calls. We score them against your playbook while you watch.