Feature log

The feature log

Everything we have shipped into the platform, newest first — written as the job each feature was hired to do.

Every entry is a Job To Be Done: when a situation comes up, what you want to happen, and so that you get a particular outcome. It is a deliberately plain format — if a feature can't be written this way, it probably wasn't worth building.

This page is updated as we ship. Dates are the month the work went live in production.

All features What's new

August 2026

Say it out loud, and it runs

A month about meeting you where you already are. A new front door lets you start from the process you run today rather than from a job you have already worked out — and plans now treat the steps only a person can do as real work orders, not just approval questions. Elsewhere: research grounded in live search; a learning loop you can read and correct; approvals on the channel each person actually watches; and machinery underneath — billing included — fast and honest enough to carry real production load.

Jobs

People's work, planned like agents' work

When: a job involves steps only a person can do — sign this, count that, call the customer
I want: the plan to model them as real work orders with an owner, a deadline and completion evidence, rather than collapsing every human touch into an approval question
so I can: run one plan that coordinates agents and people with the same rigour, instead of ending where the human work begins

Jobs

Automation that starts from how you work today

When: I know a process is eating my team's time but I have not worked out what a job for it would even look like
I want: to describe how we run it today, be asked about the parts that actually matter, and get back a proposed job I can read
so I can: automate work I already do, instead of first having to translate it into something a machine will accept

Jobs

A proposal that names your real systems and your real people

When: a proposal comes back and I have to judge whether it would survive contact with my account
I want: it to reference the systems I have actually connected, the people in my directory and my own working hours — and to name anything still missing before I accept it
so I can: find out that a step points at nothing while it is still a draft, rather than halfway through the first run

Jobs

Every proposed step marked automated, gated or manual

When: I am deciding whether a proposed job is worth accepting
I want: each step labelled as something the agents do, something they pause and ask me about first, or something that stays a person's work
so I can: see exactly how much comes off my plate — including the parts that honestly do not

Billing

An invoice that itemises what actually ran

When: a billing period closes
I want: usage rated automatically — message sends, searches and API calls priced as what they are rather than approximated as tokens — and issued as an itemised invoice with the PDF attached to the payment email
so I can: reconcile what I paid against what ran, line by line

Approvals

Approvals on the channel each person actually watches

When: a decision is waiting on someone who lives in WhatsApp, or someone who only reads SMS
I want: each person to choose the channel their approval requests arrive on, answers given there to show up everywhere else at once, and the window to expire exactly when the job says it should
so I can: have decisions reach the right pocket and be answered once, instead of chased across channels

Jobs

The learning loop, with you at the wheel

When: a job has learned something wrong, or its plan carries a hint that has gone stale
I want: to read the learnings a run actually used, edit or delete them inline, and have every hint a deleted learning motivated retire along with it
so I can: steer how a job evolves, instead of treating its memory as a black box

Platform

A workspace that reads at a glance

When: I open the platform to see where things stand
I want: jobs presented as an operations board, my agents as a team with faces and roles, services as a catalogue I can browse, and an approval queue built around answering rather than auditing
so I can: orient in seconds — what is running, who is on deck, and what is waiting on me

Agents

Research grounded in live search, with citations

When: I ask a question whose answer changes with the news
I want: Cindy to draw on live Google results and carry the citations through to the final answer — in quick questions and long research runs alike
so I can: check any claim against its source, instead of trusting training-data recall

Connections

OAuth-protected MCP servers, connected in-app

When: a tool my team relies on sits behind an MCP server that requires OAuth
I want: the authorization flow handled inside the platform when I set up the connection
so I can: give agents OAuth-protected tools without ever handling a token by hand

Connections

Beth runs your Facebook Page

When: our social presence lives on a Facebook Page
I want: to connect it — including Pages held through a Business Portfolio — and have Beth publish posts, answer comments and read page analytics
so I can: run the Page as delegated work, under the same approval rules as everything else

Jobs

Python that reads your data directly

When: a step has to work across hundreds of invoices, records or documents at once
I want: the script to ask the platform for what it needs — project files, database collections, knowledge corpora — instead of having every record staged into it one at a time
so I can: let a single step own “read all of them and reconcile”, priced like one script rather than like hundreds of model turns

Agents

Agents that compute instead of estimating

When: I ask for a count, a total, a reconciliation or a join across sources
I want: the answer produced by code that actually ran — and a plan that puts the whole batch in one step rather than one step per file
so I can: rely on the figures being arithmetic rather than a model's best impression of arithmetic

Security

A script's reach is fixed before it starts

When: code written by a model runs in the sandbox and reaches back into my data
I want: what it can see decided by a short-lived credential minted for that one run — never by anything the code itself can set or claim
so I can: be certain a script can only ever read the account and project it was launched from

Jobs

“Help me clarify” that asks about what matters

When: I have written a job definition and the platform offers to help me sharpen it
I want: the real gaps ranked up front — the ones that would genuinely change the resulting plan — and each asked about exactly once
so I can: answer four questions that change the outcome instead of a dozen generic ones

Voice

Saved voice shortcuts

When: I run the same handful of jobs over the phone every week
I want: to save them as named shortcuts that Linda recognises when I say them
so I can: start routine work with a phrase, instead of re-explaining the job every single call

Voice

Linda speaks the caller's language

When: a caller's language is not English, or my organisation's default is not English
I want: Linda to greet and answer in the right language from the first word
so I can: hand the phone to anyone on the team, or to a customer, without it feeling like a translation exercise

Voice

Shortcuts that leave a gap for the caller to fill

When: a saved shortcut needs a detail only the caller knows — a date, a name, an amount
I want: Linda to ask for the missing piece rather than guessing or failing
so I can: use one shortcut for a job that is slightly different every time

To-dos

A to-do queue for one-off work

When: a quick instruction arrives as a WhatsApp message, a phone call or an email
I want: it captured as a lightweight to-do and carried out by Agnes automatically
so I can: get small one-off requests done without the ceremony of setting up a formal job

Notifications

An in-app bell for things that are not questions

When: a job finishes something I asked for, or reports on its own progress
I want: to be told in a notification feed, acknowledged per person, rather than asked in the approval queue
so I can: keep the approval queue meaning exactly one thing — decisions that are waiting on me

Reporting

Per-job statistics: proof it is getting better

When: a job has been running for weeks and I want to know whether it is actually improving
I want: one page showing failure rate, cost per run, human time required, and the learnings it has accumulated
so I can: judge a job's trajectory at a glance, instead of reading through individual run logs

Jobs

One blocked approval no longer stalls the rest

When: part of a job is parked waiting for my decision or an external reply
I want: every independent part of the plan to keep advancing in parallel
so I can: a single pending approval never freezes work that had nothing to do with it

Jobs

Bulk work in one call instead of hundreds of round-trips

When: a job has to sweep a mailbox for attachments, pull fields out of a hundred documents, or copy a batch of files between two systems
I want: a single deterministic call that does the whole sweep in the background
so I can: pay for judgment rather than for an agent clicking through the same operation two hundred times

Platform

Model routing you control, and can see

When: I care what each part of the platform costs and where it runs
I want: every model call routed through a provider I choose, with the model that actually ran shown on the message itself
so I can: treat cost and compliance as settings I own, rather than as surprises on an invoice

Automation

Triggers that explain themselves

When: I want something outside the platform to start a job — a form, an inbox, a webhook, a schedule
I want: to describe it in plain language, follow a guided go-live checklist, and see exactly how the incoming data reaches the job's inputs
so I can: wire up automation without first learning the difference between four ingress types

Administration

Open the product as one of your users

When: someone on my team reports a problem I cannot reproduce
I want: to open the platform as them — without their password, clearly labelled, and time-limited
so I can: see exactly what they see instead of asking them for screenshots

Platform

One governed home for project files

When: agents read, write and move files during a job
I want: all of it to run through a single governed file store with one set of rules
so I can: know where a job's working files are, and that nothing reached them by a side door

Security

Every internal surface authenticated, every write attributed

When: the platform is exposed to the public internet across many servers
I want: session identity derived server-side, internal surfaces authenticated, and spend limits counted across the whole fleet rather than per server
so I can: rely on the access controls and budget caps actually holding, instead of holding on one machine at a time

Performance

Built for a busy day

When: many jobs, chats and viewers are active at the same time
I want: incremental event polling, chat streams that resume across servers, indexed hot paths and cached static reads
so I can: have a busy afternoon distribute across the platform instead of cascading into timeouts

July 2026

Connect what you already run

The month the platform stopped being an island. Microsoft 365, accounting, any IMAP mailbox, your own internal APIs — plus the white-label layer that lets a partner sell the whole thing under their own name.

Connections

Microsoft 365, at parity with Google

When: my team runs on Outlook, OneDrive, Teams and Outlook Calendar
I want: to connect those accounts and have agents read and send mail, manage files, post to channels and book meetings
so I can: get the benefit of agentic automation without being locked into one vendor's ecosystem

Connections

Accounting through Bokio

When: my business keeps its books in Bokio
I want: agents to read and write journal entries, invoices, customers and receipts directly
so I can: let accounting tasks finish inside a job, with no manual re-entry at the end

Connections

Any mailbox, over plain IMAP and SMTP

When: my mail is not on Gmail or Microsoft
I want: to connect it with standard IMAP/SMTP credentials
so I can: run email work from whatever mailbox my business actually uses

Connections

Several accounts of the same kind, told apart by name

When: we have three Gmail inboxes, two Dropbox accounts and more than one OneDrive
I want: to connect them all under distinct names that agents can target explicitly
so I can: be certain a job used the invoices mailbox and not the support one

Service providers

Sell the platform as your own

When: I want to deliver agnticr to my customers under my own brand
I want: a hierarchical service-provider layer with custom domains, inherited branding, host-scoped sign-in and my own agent naming
so I can: hand each customer a platform that feels entirely mine, with no per-customer engineering

Agent services

People Coordination: get a meeting agreed without the email chain

When: I need to agree a time with several people outside my company
I want: a professional slot-options email with one-click accept, a response matrix showing who agreed to what, and a calendar where I can step in and confirm
so I can: close a coordination round without a fifteen-message thread, and without anyone having to create an account

Agent services

Bonnie, the facility scheduler

When: I plan a week of slots for teams, rooms or resources
I want: to pin the fixed bookings, auto-plan around them, then drag to adjust with push, swap and undo — and share or export the result
so I can: run weekly scheduling as a guided, reversible workflow rather than a manual grid puzzle

Agents

Wendy reads the live web, safely

When: a job needs something that only exists on a public web page — a price, a listing, a timetable
I want: a strictly read-only browser worker that fetches the real page
so I can: ground work in what the page says today, without opening a door into our own network

Jobs

Real Python, in a locked-down sandbox

When: a deliverable needs actual computation — a rich Excel report, a multi-page PDF, a chart, a large dataset
I want: the agent to run real code in an isolated sandbox with the libraries already installed
so I can: receive a file I can open and use, produced by arithmetic rather than by a model's best guess

Jobs

A review that repairs the plan

When: a run finishes and something in the plan clearly needs to change
I want: the review to propose the repair, show me the before/after of the milestone text, and record why
so I can: watch the plan improve between runs, with every change attributable

Jobs

Jobs start from what worked last time

When: a recurring job has already solved a problem once
I want: the next run briefed on the tool sequence that worked and the pitfalls it hit
so I can: get a tenth run that is measurably faster and cheaper than the first

Jobs

Batches that fan out — and partially succeed

When: a job has to work through hundreds of independent items
I want: each item handled by its own isolated worker, with successes counted as done and only the failures flagged
so I can: process mailbox-scale volumes in one run, and never lose a batch to one bad record

Jobs

A binding contract before the first run

When: I describe a job in a few sentences of plain language
I want: the planner to turn it into concrete success criteria, scope boundaries and a cost and time estimate I can read
so I can: start every run with the agent and me holding the same definition of done

Billing

A hard monthly spend cap

When: I am responsible for what this costs
I want: a monthly ceiling enforced across every channel — chat, jobs, voice, triggers, WhatsApp
so I can: sleep through the night without checking a usage dashboard

Platform

Swedish and English, everywhere

When: my organisation works across Swedish- and English-speaking markets
I want: every surface — the app, jobs, voice, transactional email, approvals and outbound messages — in the reader's own language
so I can: give a Swedish colleague and an English customer the same product, each in their own language

Governance

Scoped memberships and explicit roles

When: several people share an account and some of them should not be able to do everything
I want: per-account memberships with explicit roles, rather than platform-wide admin
so I can: audit who can do what, account by account

Approvals

Irreversible actions stop for a human

When: an agent is about to do something it cannot undo — send a message, delete a record, write to an external system
I want: that action gated on an explicit approval before it fires
so I can: never learn about a consequential action after the fact

Governance

An audit log of what people and agents did

When: I need to investigate an incident, a billing anomaly or an unexpected change
I want: a structured, searchable trail of user and agent actions
so I can: answer the question without trawling server logs

Platform

The parts that matter work on a phone

When: an approval reaches me while I am standing in a school pickup line
I want: approvals, job status and external conversations to work properly on a narrow screen
so I can: decide in seconds without going back to a laptop

June 2026

The department grows up

The biggest build month so far: jobs got a memory, plans started evolving themselves, cost became something you can see and steer, and the platform opened up to external tooling over MCP.

Jobs

Plans that evolve themselves, with provenance

When: each run of a job teaches the platform something about what actually works
I want: the plan to update itself between runs as a new, versioned revision that records why it changed — and bad lessons to be retired
so I can: have a job improve without me hand-editing it, and still be able to see exactly what changed and on what evidence

Jobs

A per-job workspace — the digital worker

When: a job runs every week and keeps needing the same context
I want: it to have its own persistent workspace: files, a database, a knowledge base and a memory
so I can: get a dedicated worker that accumulates context, rather than a stranger starting from zero every Monday

Agents

Memory that curates itself

When: I tell an agent something durable about how we work
I want: it remembered and recalled automatically in later sessions
so I can: stop repeating myself, without maintaining a memory file

Platform

The right model for each step, escalating on failure

When: a job has both mechanical steps and steps that need real judgment
I want: each step routed to a model sized to its difficulty, escalating automatically when one fails
so I can: get quality where it matters without overpaying everywhere else

Jobs

Receipts the judge cannot argue with

When: the platform decides whether a run actually succeeded
I want: the verdict grounded in non-fabricable receipts of every external side effect — message ids, record ids, delivery confirmations
so I can: trust a green run because of what it can prove, not because of what it claims

Reporting

Time saved, per run and cumulative

When: I want to know whether this is worth what it costs
I want: each completed run to show the machine time spent and the manual effort it replaced, compounding across runs
so I can: measure the value of the automation without building a tracker for it

Jobs

Delegated sub-agents, running in parallel

When: one step of a job splits into several independent pieces of work
I want: them handed to fresh isolated workers that run at the same time and return only their results
so I can: finish faster, and pay in proportion to the reasoning that actually mattered

Automation

Triggers, made human

When: I want a job to start from an incoming email, a form, a webhook or a schedule
I want: plain-language setup, readable URLs instead of raw identifiers, a live readiness checklist and a way to test against a sample payload before going live
so I can: wire up automation in minutes without writing an expression

Data

A shared document database for job data

When: work produces structured data that later runs will need
I want: it stored in a shared project database, with a spreadsheet-style editor for people who do not think in schemas
so I can: reuse results across runs instead of leaving them trapped in one execution

Reporting

Custom Canvas: describe a dashboard, get a dashboard

When: I want a view over a job's real data
I want: to describe it in one prompt and get a chartable, read-only, network-less widget
so I can: get insight without building or wiring anything, and with no way for the view to change my data

Agents

Teddy produces real files

When: a job has to deliver a report, a deck or a spreadsheet
I want: a genuine .docx, .pptx, .xlsx or .pdf, not a text approximation of one
so I can: forward it to a colleague without a conversion step in between

Platform

Reach your account from your own tooling, over MCP

When: I already work in Claude Code or another external runtime
I want: a secure per-account MCP endpoint that exposes what a human operator can reach — jobs, chat, approvals, executions, data
so I can: drive and diagnose the platform from outside it, without sharing a raw API key

Platform

Sign in with Google or Microsoft

When: my organisation already has an identity provider
I want: to sign in with the account I have, rather than managing another password
so I can: onboard a colleague in one click

Governance

Project guidelines: standing instructions every agent honours

When: I have house rules — tone of voice, audience, do's and don'ts, approval rules, glossary
I want: to write them once for the project, like a briefing note
so I can: have every agent, in every chat and every job, follow them without me repeating myself

Performance

Large results stay out of the conversation

When: a tool returns a huge result — a record set, a long web page, a mailbox listing
I want: it bounded and stashed rather than flooded into the shared context, retrievable on demand
so I can: pay for the tokens that drove reasoning, not for the same payload re-billed on every turn

Jobs

Every loop has a hard exit

When: an agent job runs for hours
I want: explicit stop rules — schedule limits, convergence checks, per-run token budgets, per-task success criteria
so I can: have a runaway loop fail fast and honestly instead of quietly exhausting a budget

Security

A web reader that refuses to be hijacked

When: agents fetch pages from the open internet
I want: internal network targets refused, instructions embedded in page bodies ignored, and crawling bounded
so I can: read the web without the web being able to redirect what my agents do

Administration

Manage your own team

When: someone joins or leaves
I want: to invite by email, trigger a password reset, see who last signed in and mail users directly from the admin panel
so I can: handle it myself, without filing a support request

May 2026

Humans in the loop, understood

Approvals stopped being a form to fill in. When you reply to a gate, the platform now reads what you meant and folds it into the run — and every finished execution gets reviewed.

Approvals

Your reply is read, not parsed

When: a job pauses for my decision and I answer in my own words
I want: the platform to interpret what I meant — approval, refusal, or a change of direction — and act on it
so I can: answer a gate the way I would answer a colleague, instead of matching a required keyword

Approvals

Guidance becomes part of the plan

When: I approve a step but add a note about how to do the next one
I want: that guidance persisted as a real step in the execution
so I can: have my instruction survive the rest of the run, not just the next turn

Jobs

Every execution gets reviewed

When: a run finishes, well or badly
I want: a structured review of what happened and what it means for the next run
so I can: learn something from each run rather than just filing it

Agents

Knowledge skills and memory workflows

When: my organisation has its own material an agent should know
I want: it available to agents as retrievable knowledge, with memory workflows that keep it current
so I can: get answers grounded in our documents rather than in general knowledge

Jobs

Retries grouped in the execution timeline

When: a step needed three attempts before it worked
I want: those attempts grouped under one step rather than listed as three separate events
so I can: read the timeline as the story of the run, not as a stack trace

Platform

Running on Kubernetes

When: the platform has to survive deploys, restarts and traffic spikes
I want: it running on a real orchestrated cluster
so I can: have long-running jobs continue across a deploy instead of dying with the server

April 2026

Gates, triggers and budgets

The three things that turn an interesting demo into something you can leave running: a real approval protocol, a way for the outside world to start work, and a cap on what it can spend.

Approvals

Approval gates as a system, not a prompt

When: a plan contains steps I must sign off before they happen
I want: checkpoints, conditional gating and input requests declared in the plan and enforced by the engine, with every response logged
so I can: rely on the gate holding because the engine enforces it, not because the model remembered to ask

Automation

Jobs started from outside the platform

When: the event that should start a job happens somewhere else
I want: external triggers that can kick a job off without anyone opening the app
so I can: have work begin the moment its trigger arrives

Billing

Budgets per account

When: I run more than one account on the platform
I want: a spending budget attached to each one
so I can: give a team autonomy without giving it an unlimited card

Agent services

An agent that corresponds with people outside your company

When: a job needs a real back-and-forth with a supplier, a customer or a contractor
I want: an agent that can hold that conversation over email and report back
so I can: have chasing and following up happen without me in the middle of it

Agents

A document specialist

When: the deliverable is a document rather than an answer
I want: a specialist whose whole job is producing it properly
so I can: get output that looks like it came from a colleague

Jobs

System playbooks

When: a kind of work recurs across many accounts
I want: proven procedures available to the planner by name
so I can: start from a known-good approach instead of a blank page

Jobs

Author a job without leaving the conversation

When: I realise mid-chat that this should be a recurring job
I want: to create and shape it inline, right there
so I can: go from an idea to a scheduled job in one conversation

March 2026

From chat to a plan you can approve

The month the core idea landed: describe the work in plain language, get back a milestone plan you can read, change and approve — and answer its questions from WhatsApp.

Jobs

Job planning: a plan you can actually read

When: I describe a recurring piece of work in a few sentences
I want: a validated milestone plan back — one I can review, edit and regenerate before anything runs
so I can: approve the approach before committing to it, not after

Jobs

Watch the plan execute, milestone by milestone

When: a job is running
I want: to see which milestone it is on, what it has done and where it is waiting
so I can: know the state of the work without asking anyone

Approvals

Approvals over WhatsApp

When: a job needs a decision and I am not at a desk
I want: the question delivered to WhatsApp, with enough context to answer safely
so I can: unblock a run from my phone in under a minute

Connections

Connect your own internal APIs

When: the system that matters most to my business is one we built ourselves
I want: to describe it to the platform and have agents call it under governed, allowlisted access
so I can: automate around our own systems, not only around off-the-shelf ones

Jobs

The plan canvas

When: a plan has branches, dependencies and gates
I want: to see it laid out visually rather than as a list
so I can: understand the shape of the work at a glance

Data

A document database scoped to your account

When: jobs need somewhere structured to keep their data
I want: a document database isolated to my account and my projects
so I can: keep our data separate from everyone else's, by construction

Billing

Token usage reporting

When: I want to know what a run cost
I want: usage recorded and reported per run
so I can: see the price of a job next to its result

Platform

An in-product help portal, in your language

When: I need to know how something works while I am using it
I want: help available inside the product, in Swedish or English
so I can: answer my own question without leaving the page

February 2026

Foundations

The first month of the platform proper: accounts, projects, the jobs workspace, and the agent roster that would end up doing all the work.

Platform

Accounts, projects and scoped sessions

When: more than one organisation, team or line of work shares the platform
I want: a structure where accounts hold projects, and every session, file and conversation is scoped to one of them
so I can: keep separate work separate, from the very first day

Jobs

The jobs workspace

When: I have several jobs at different stages
I want: a workspace that shows them as cards with a visual timeline of each run
so I can: see what is planned, running and finished in one view

Platform

A project home page

When: I open a project after a few days away
I want: an overview of what lives in it and what has happened recently
so I can: pick the thread back up immediately

Agents

The agent team, made visible

When: work is done by a roster of specialists rather than one assistant
I want: a page that shows who they are and how they hand work to each other
so I can: understand who did what, rather than treating it as one opaque system

Data

A project filesystem

When: agents produce and consume files
I want: a real filesystem per project that both the agents and I can see
so I can: find a job's output where I would expect it to be

Data

Knowledge corpora you manage yourself

When: agents should answer from our documents
I want: to upload and manage knowledge corpora, and see clearly when one fails
so I can: control what the agents know without a developer in the loop

Want this running in your business?

Join the waitlist and we'll show you what these features look like on your own work.

Join the waitlist See every feature