This week agents became roles with a page of their own, finance got a Plan dashboard with a forecast, Balio's side was cut back to a few pages, and every change now reaches production through review and a staging step.

Highlights

• Agents: under Office, every agent is now a role with a job title, such as Customer Support Specialist or AR Specialist. One card per role, and Hire adds one. Agents outside the customer department are paused for now. (Orhan)
• Agent page: open a card for the tabs Overview, Operations, Tests, Setup, Integrations and Activity. Mode sets how much the agent does alone: Draft, Controlled, Auto or Full. (Orhan)
• Tests: the agent answers your closed tickets of the last 90 days. Nothing is sent. Each answer sits beside your team's real reply and gets a score. (Orhan)

• Inbox: weekly reports no longer land here. A note nobody opened in 14 days expires, and so does a drafted reply once a person has answered the customer. (Orhan)
• Plan: a new dashboard with Forecast and Customers. It shows actual figures, then the forecast to the end of next year. Add your own changes as drafts, and keep Notes and key dates. (Sydney)

• Revenue and P&L: every chart has a Chart and Table switch. Click a chart to filter, and a Filtered by row names what is set. (Sydney)
• Usage: under Product, active users, active accounts and feature adoption from your analytics tool. (Orhan)
• Workspaces: everyone now lands here. It replaces Home, Tools, Agents, Planning, Experiments and the command palette. Each workspace opens to Overview, Integrations, Automations, Members, Usage and Controls. (Orhan)

• Cost: one page for agent spend across all workspaces, by workspace, agent and model. (Orhan)

Everything shipped this week

Agents and automation

• Agents as roles: each agent is a role with a job title, one card per hired role, hired by role, and the old mythological names are gone (Orhan)
• Customer roles: the customer department is now eight roles, and every hire is called by its title (Orhan)
• Role card: the card of a role was redesigned, the summary row above the cards was removed, and roles are grouped under their teams (Orhan)
• Agent page: one page for every agent with six tabs, a readable work queue, tools by category, the four modes Draft, Controlled, Auto and Full, and settings the engine really enforces (Orhan)
• Tests tab: an agent can be tried on the workspace's own closed tickets before it is hired, with a score and nothing sent (Orhan)
• Autonomy by role: one autonomy setting per role on its page, in place of one per duty (Orhan)
• 3D office: the office is now a 3D floor with one desk per role, it is the page a workspace opens on, and a desk opens the role's page (Orhan)
• Simpler workspace menu: Today is Jarvis's chat only, and Team became Agents under Office (Orhan)
• Customer department first: every agent outside the customer department is paused in every workspace until the customer team is designed, and nothing else about them changed (Orhan)
• Evolution loop: each night the engine may try one tested improvement per duty and shows a card only when the result is clearly better. Built with a nightly spending cap, and the switch-on for AnnounceKit waited for the GM's OK (Orhan)
• Onboarding run: the prove and shadow steps of a new workspace's onboarding now open and finish by themselves (Orhan)
• Product learner: a new learner reads product events, releases and tickets to describe features, retention and how often the company ships (Orhan)
• Organization learner: a new learner describes the team's load per department from seats, desk work and meetings, with seat details visible to the GM only (Orhan)
• Market signals and pricing: the market learner now reads reach, hiring and news about the market each month, and a pricing suggestion is built from past price changes and competitors' prices (Orhan)
• Learner fixes: learners keep the better of two answers instead of the newest, the Slack learner counts a stated rule as a finding and says why a run found little, and the customer learner's slowest step was made fast (Orhan)
• Company profile freshness: each profile section carries when it was observed and until when it holds, agents are told when a section is stale, and web evidence is cached per workspace (Orhan)
• Engine run accounting: a run reserves its budget before it starts and settles after, upkeep spend counts in the monthly limit, and a failed run is no longer read as an empty one (Orhan)

Inbox and approvals

• Inbox by role: the Inbox names each item by the role that made it and can be narrowed to one role (Orhan)
• Weekly reports out of the Inbox: reports stay in the run's log, and at-risk accounts arrive as one weekly digest, with a separate card only for a VIP or critical account somebody owns (Orhan)
• Notes expire: a note nobody read in 14 days expires, except security, VIP, refund and legal ones and those addressed to a person (Orhan)
• Answered drafts expire: a drafted reply leaves the Inbox when the thread shows a person already answered the customer (Orhan)
• VIP hand-offs in Slack: a VIP hand-off from any agent now also reaches its owner in Slack, tied to its Inbox item (Orhan)
• Help center proposals: a proposed help article change is an Inbox task plus a chat post with Accept and Decline, only in a channel verified as the company's own (Orhan)
• Decision log: explicit decisions found in stored Slack and meeting notes are logged and come back to the Inbox for review when they are due (Orhan)
• Pause all: pausing every agent now happens in one step, and the app asks before discarding drafts that are waiting (Orhan)
• AnnounceKit Inbox clean-up: closed, answered and superseded support items were expired from AnnounceKit's waiting Inbox, counting answers written from the shared admin account (Orhan)

Support desk

• Last look before sending: after the support audit of 4 Nov, a customer reply is read against the thread once more before it leaves, and a held draft goes out only on a person's decision (Orhan)
• Holding message: AnnounceKit's "we are on it" message was switched off after the audit, then switched on again behind guards. It stops when a person is on the case, and its reminder note closes itself (Orhan)
• Fact check on a second model: the check before an automatic reply can run on a different model family than the one that wrote the reply (Orhan)
• AnnounceKit knowledge base: the desk now reads AnnounceKit's own Knowledgebase articles as help documents, and the help center keeper posts its proposals to the support channel (Orhan)

Dashboards

• Plan dashboard: a new dashboard in the sidebar shows the forecast beside the actuals to the end of next year, with forms where the GM enters changes and sees them in the graphs at once (Sydney)
• Plan, Customers page: the forecast per customer and the renewals due, where opening a customer lets you change it (Sydney)
• Drafts, notes and key dates: Plan changes can stay drafts, notes and key dates are saved on the dashboards and drawn on the charts, with alerts and an anonymous comparison with the portfolio (Sydney)
• Chart and Table switch: every Revenue and P&L chart can be read as a table (Sydney)
• Filters: quick date ranges and Clear, a Filtered by row that names every filter set by a click, and the light theme by default (Sydney)
• Cash flow: the Revenue cash flow page shows how customers paid, by card or by bank (Sydney)
• P&L: the P&L now ends at EBITDA, and lines removed in the warehouse disappear from Argus too (Sydney)
• Open month left out: churn lists, cohort views and customer drill-downs no longer count the month still in progress, which made them look worse than they were (Sydney)
• Revenue charts: one axis per chart, compared periods as a dashed line, cohorts by calendar month, and the Needs attention list removed from the overview (Sydney)
• Dashboard fixes: Past due lost the chosen date range and now keeps it, a carried-over cohort filter on Churn is now visible, and several layout faults were fixed (Sydney)
• Plain words: the wording across Revenue, P&L and Plan was rewritten in plain language (Sydney)
• Product, Usage: a new page with active users, active accounts and feature adoption from the analytics tool's events (Orhan)
• Business review: new Forecast, Capacity and Risks blocks: MRR, cash and churn at 30, 60 and 90 days marked reliable or not, each department's load against its capacity, and a monthly list of risks (Orhan)
• Company goals: an admin sets goals, and each week Argus shows the pace, what drives it and a one-line explanation (Orhan)
• Portfolio benchmark: each quarter a workspace sees its own position against anonymous portfolio figures, published only when enough workspaces report (Orhan)
• Look and feel: a new Argus logo and colours, one icon per page, chart legends as toggles, and the app's own dropdowns everywhere (Orhan)
• App review fixes: about thirty small faults from a review of the app were fixed, such as money shown in the workspace's currency, names in place of internal keys, and one main button per card (Orhan)

Data and connections

• Ahrefs: a new read-only connection that gives the market learner keyword demand, competitors' search share and visibility in AI answers, within a monthly cap (Orhan)
• Outbound: Clay and Instantly campaigns with their full cost arrived as new Marketing pages, visible to Balio's admins only for now (Tiaan, Orhan)
• Instantly: Instantly is now an Argus integration, and Outbound reads its Clay, Instantly and Claude keys from Integrations (Tiaan)
• Data Report: a gap a person has accepted leaves the list and is shown apart (Orhan)
• Switched-off connectors: a connector its owner turned off is no longer reported as a broken tool (Orhan)
• Data quality score: the score now compares like with like across sources and splits missing contacts by cause, so it stops penalising rows that cannot be linked (Orhan)
• One definition of paying: every part of the engine now counts a paying customer the same way, on the company's own day (Orhan)
• Sync and retention fixes: a source with no data is retried once per period instead of being left dead, the nightly trim keeps records that others still point to, and the Slack scan keeps each channel's newest weeks (Orhan)

Access and security

• Security hardening: warehouse settings can be changed by Balio only, a tool connection can be finished only by the person who started it, retention settings stick, and a suspended seat loses See as (Sydney)
• Jarvis settings: every seat keeps the Jarvis tab, and people who are not admins see only their own sections (Orhan)
• Privacy checks: the check that personal data is covered by erasure now also looks at phone numbers, photos and links, and erasure clears photos and invoice links (Orhan)

Balio's side

• Workspaces: Balio's dashboard was cut back to the Workspaces list, with real tool logos, an agent cost column and no warning chips. Home, Tools, Agents and Planning are gone (Orhan)
• Workspace page: each workspace has the tabs Overview, Integrations, Automations, Members, Usage and Controls, and Overview shows what the agents changed per department (Orhan)
• Cost: a new page with agent spend across every workspace, by workspace, agent and model. A workspace's Usage tab no longer shows cost (Orhan)
• Experiments removed: agents' trials were taken out of the product, and the old addresses land on Workspaces (Orhan)
• Admin, Controls and Security: Admin and Controls got tabs and calmer tables, and Security shows its chart-led score dashboard again (Orhan)
• Evals: Compare shows error bars and says when a difference is only noise, Calibrate takes a second labeller with export and import and has a plain-language guide, and AnnounceKit's model key now runs the calibration judges (Orhan)

Platform and engineering

• Every change through review: all changes now land through a pull request that a code owner approves, with a written guide and a pre-flight check for teammates (Orhan)
• Staging before production: the working branch now deploys to a separate staging copy, and only the main branch deploys to production (Marcelo, Orhan)
• Preview per change: each pull request can get its own preview of the dashboard, and its own backend when asked for, removed again when it closes (Marcelo, Orhan)
• Staging secrets apart: staging reads only its own set of keys, never production's (Marcelo)
• Merge queue and test cost: the heavy test suites run once per merged change, in a queue that tests changes one after another, which cut the test bill (Orhan)
• Faster, steadier tests: the test run was sped up, and tests that failed only on certain days of the month or times of day were fixed (Orhan)
• Fair scheduling: the engine's hourly upkeep jobs now take turns and are bounded, so one slow job cannot starve the others (Orhan)
• Diagnose tools: new read-only reports for the support desk, the waiting Inbox, data quality, the learners and the sync, used to prove this week's fixes on real data (Orhan)