ACTAVA Platform Release Notes v9 (October 2026)

Copy rich textContent selected. Press Cmd+C to copy.

ACTAVA Platform Release Notes KORA v9 (October 2026)

Release date: October 31, 2026

Release at a glance: A trust architecture built on seven principles, an organization-wide Emergency Brake, cross-provider model resilience under your consent, human review between workflow steps, a Scheduled hub, deadline-driven batches, a new Agent Dashboard, voice agent experiments, Task Map, and more than 70 further improvements across Agent Studio, Pipeline Builder, the Partner API, Voice, evaluation, security, and reliability.

Governed Autonomy: Run Agents Unattended. Stay in Command.

Enterprises are handing AI agents real work: scheduled jobs that run overnight, batches of thousands of items, voice agents that answer the phone, and multi-step workflows that write to production systems. The more autonomy those agents get, the more control the organization needs over them. A capable model is not enough on its own.

ACTAVA V9 is built on one principle: separate the supply of intelligence from the authority over it. Models do the reasoning. The organization decides what they can reach, which actions need a person, when they must stop, and how every outcome is verified. In V9 those controls sit outside the model, so no model can bypass them.

V8 opened the platform to the teams who build on it. V9 makes sure the organization running those agents can always step in: an organization-wide Emergency Brake, resilience across model providers that the organization consents to, human review between workflow steps, independent verification of every agent, and a trust architecture an auditor can check.

V9 is also a measure of how we work. Every month, our engineering team ships a full release of capabilities our customers asked for, built with the same discipline: recorded architecture decisions, automated tests on every change, verification in staging, and security review. This month we went further and audited the entire platform against an insider-risk standard before inviting an external penetration test.

ACTAVA KORA covers Build, Test, and Learn, and ACTAVA Compliance covers Guide. V9 gives both the controls to run agents unattended with confidence.

Book a demo to explore V9 with your own workflows.

V9 at a Glance

Six headline capabilities

1. Trust Architecture. Controls, evidence, and verification that sit outside the model, mapped to seven principles for containing powerful AI.

2. Emergency Brake and Durable Stop. Stop or hold every run in an organization, or on one agent, until a person releases it.

3. Multi-Model Resilience. Fallback across providers and managed model retirement, both under the organization's consent.

4. Human Review at Every Step. Decision Layer approval groups on Agent Workspace steps and tool calls.

5. Autonomous Operations. The Scheduled hub, a deadline-driven Agent Batch API, and a new Agent Dashboard.

6. Evidence-Driven Agent Quality. Voice agent experiments with statistical comparison, Task Map, and server-side evaluation.

Plus the Capabilities hub, ACTAVA-built agents with starter checks, a much larger Pipeline Builder, an expanded Partner API and MCP surface, per-organization phone numbers, an updated model catalog, and 73 improvements in the complete feature list.

Engineering by the numbers

One month of engineering, in numbers.

• 2,300+ pull requests merged across 9 platform services, each one reviewed and tested before it lands.

• 455 architecture decisions recorded. Every structural choice is written down with its reasoning and the alternatives considered, so the platform stays explainable as it grows.

• 18 release builds and 2 production promotions. Every build is verified in staging before it reaches customers.

• A full security hardening pass across the API, agent runtime, router, and sandbox, followed by an external grey-box penetration test.

Availability

• Trust Architecture (model diversity, run integrity, audit export) · Security, compliance, and admins · Available; some controls roll out per environment

• Emergency Brake and Durable Stop · Org admins and super admins · Available

• Multi-Model Resilience · Org admins, all runs · Available, organization opt-in

• Updated model catalog (GPT-6.1 Sol, Sonnet 5.5, Haiku 5.5) · Everyone · Available

• Decision Layer and Workspace review gates · Admins, reviewers, and developers · Beta

• Scheduled hub · Every role · Beta

• Agent Batch API updates · Partners and API users · Available

• New Agent Dashboard · Developers and admins · Preview

• Voice agent experiments and comparisons · Developers and admins · Available where voice benchmarking is enabled

• Task Map · Developers and admins · Beta

• Capabilities hub · Every role · Available in organizations on the new skills experience

• ACTAVA-built agents with overview and starter checks · Developers and admins · Available

• Pipeline Builder expansion · Developers and admins · Beta

• Partner API and KORA MCP additions · Partners and API users · Available

• Per-organization phone numbers · Org admins · Provisioned by ACTAVA

• Microsoft 365 personal connections and per-service switches · Everyone · Available

• Text to Speech and ElevenLabs tools · Developers · Available

Headline Capabilities

1. Trust Architecture

Trust the system, not the model.

Frontier models are now more capable than the software around them, but their behavior cannot be traced to a code path. ACTAVA treats every model like a powerful insider: useful, and kept inside boundaries it cannot change. V9 is the release where we audited the whole platform against that standard, starting with an internal insider-risk controls review on October 10 and continuing into an external grey-box penetration test that began October 26.

Here is how V9 delivers on each of the seven principles.

• Model diversity. No single model is the only dependency, and none checks its own work. Every checker, judge, and self-check runs on a different model from the agent it grades, preferring a different provider. A same-model pin is swapped at run time and recorded, and saving a new judge pinned to the agent's own model is refused. Backup models come from a different provider.

• Observe everything. Every meaningful action leaves evidence a person can read, without asking the model. A nightly keyed integrity digest of each organization's runs means a changed or missing record can be detected. Finished turns are frozen, and each run records its code version, agent card, the configuration it received, and how it was launched. Runs are soft-deleted, never silently erased.

• Verifiability. Test the whole system, including failures and edge cases. Voice agent experiments with statistical comparison, Task Map, server-side and auto-run evaluation, and starter checks on every ACTAVA-built agent.

• Independent controls. The organization decides what an agent can reach and do. Role-scoped API keys, Decision Layer review gates, consent for model changes, per-agent connector identity, a sandbox floor the agent cannot remove, and no tenant-launched stdio MCP servers.

• Independent auditability. Validation does not depend on the intelligence being validated. An auditor export whose run evidence can be checked against the integrity digests. Only the agent runtime can write run records. The compliance attestation claims only what the platform enforces, and a test fails on any unbacked claim.

• Containment. An authorized person can always pause or stop an agent mid-task. The Emergency Brake and a durable Stop that holds through resumes, child runs, and background jobs.

• Incident disclosure. When something fails, the people affected find out, with the cause. A public status page driven by direct checks, SLA measurement with a live status view, and notices that name the cause of a failed run, a paused schedule, or a model retirement.

Already in place from earlier releases: mandatory turn gates, approval-gated tools, deterministic safety rules, Slack channel scoping, the web domain allowlist, Voice Call Replay, citations on every answer, and the human-in-the-loop decision record.

2. Emergency Brake and Durable Stop

Stop everything, or hold it, until a person says go.

When an agent misbehaves in production, two things matter: how fast you can stop it, and whether it stays stopped.

• Organization or agent scope. An org admin (in their own organization) or a super admin engages, escalates, or releases a brake for the whole organization or for one agent, on one version or all of them. Organization scope covers agents, ACTAVA, evaluations, benchmarks, and Synthesis Studio.

• Stop or Hold. Stop cancels in-scope runs, cascades to their child runs, withdraws pending approvals, and aborts live Agent Workspace runs. Hold pauses running turns so they can continue after release; decisions made during a hold are recorded and handed to the run on release.

• Checked before every start. Every run start and resume checks the brake first and fails closed. Batches pause rather than fail, and a refused start is reported as a retryable policy block.

• A human decision. API keys, the scheduler, and pipelines cannot engage or release a brake, and an org admin cannot release a brake a super admin engaged. Every action is audited and announced.

• A Stop that sticks. A Stop is written durably, applies only to the turn it was meant for, stops child runs and background jobs, and holds however many times a paused run is resumed. Gated tool calls stopped this way are never dispatched later.

• Security & Audit. Emergency Brake and Audit export are standalone pages in a new Security & Audit sidebar group.

3. Multi-Model Resilience

Keep work running when a provider fails, on your terms.

Model providers have outages and retire models. V9 lets agents survive both, while the decision to switch models stays with the organization.

• Cross-provider fallback. When the primary model fails, a run can continue on a backup model from a different provider. Background and scheduled runs retry transient failures, and a circuit breaker stops repeated calls to a broken configuration.

• Bounded calls. A provider error inside an open stream counts as a provider failure, and every model call is bounded in time.

• Managed retirement. Runs pinned to a retired model use its declared successor, and are billed and recorded as the model that actually ran. Users see retirement notices in chat and the Builder, and each person gets one email per retired model.

• Organization consent. Two independent switches in model access: Automatic model fallback (off by default) and Automatic model upgrades (held until an admin approves).

• Clear failure reporting. Model failures finalize the run and raise a run_failed notice, deferred while a retry is pending. A backup that refuses counts as a failed fallback.

• Updated catalog. Regular moves to GPT-6.1 Sol, Sonnet 5.5 and Haiku 5.5 join the catalog, and legacy pins offer one-click Upgrade to the current tier model.

4. Human Review at Every Step (beta)

Put a person between the steps that matter.

The Decision Layer extends human-in-the-loop review beyond Approve and Reject, and brings approval groups to multi-agent workflows.

• Workspace review gates. An Agent Workspace step can carry a review gate. The run holds after that step, an approval group decides, and the next step reads only approved files.

• Tool-call review. Gated tool calls are reviewed in full and recorded per call. Chat shows an inline approval card, says which group the run is waiting on, and reflects a decision made elsewhere.

• Decisions with context. A run stopped by an approval tells its owner which decision stopped it, who made it, and what happens next. Decision pages say what is being decided, confirm a reject, and show a decided request as read-only.

• Rework that holds. Return for rework re-sends the planned range and holds the approval for the redo, and a re-asked decision says what the agent redid.

• Safe by construction. Approval groups travel with exported agents. A gate nobody can answer never runs, test runs are refused while an approval group is empty, and a new message cannot get past a hand-off held for approval.

• Rollout. Enabled per organization with the decision_layer flag, which also gates Run Governance. Workspace review gates are enabled separately and support accept and reject in this phase.

5. Autonomous Operations

Agents that run on schedule, at volume, and in plain view.

Scheduled hub (beta). One place for every scheduled agent, and one way to schedule it.

• One schedule editor. Once, hourly, daily, weekdays, weekly, monthly (day N, last day, first or last business day), or cron, with a required time zone. An agent can have several named schedules, and the server computes the next fire times.

• A page for every role. Scheduled sits under My work for every role, with an Organization tab for admins, search, status filters, and schedule details.

• Schedule in conversation. ACTAVA can create a schedule in a short conversation, or build an agent and schedule it in the same conversation. A schedule on a draft starts when the agent goes live, and Tell me when it matters controls when a finished run sends a notice.

• Built for unattended runs. A scheduled run never waits on a question nobody will answer. A schedule pauses itself after three runs that could not run for a fixable reason and says why, and notices name the missing connector.

• Files that carry over. Schedules can use a durable project, so scheduled runs keep that project's files from one run to the next.

Agent Batch API. Deadlines and quotas that follow the work.

• Deadlines from the work. A batch's target deadline is derived from its unfinished work (at least 24 hours, sized at 2,100 items per hour per organization), with a hard expiry between 48 and 168 hours.

• Cost and capacity. Batch cost in OU, served_ratio, a batch.completed webhook, and quota-threshold webhooks.

• Pause, don't fail. Over-quota batches pause instead of failing, and every billable start follows one quota rule (403 OU_QUOTA_EXCEEDED).

• Item-level accounting. Agent versions can declare a deliverable tool, and requests accept items (1 to 1,000 ids) for per-turn item counts.

• Honest run reporting. API Runs reports partial completions, shed items, agent names, and why a run failed. Submissions lost before their first save are recovered.

New Agent Dashboard (preview). Every agent's health at a glance.

• An agents table with recent runs, day click-through, charts, favorites, prompt search, and failed runs that show an error code and a redacted message.

• A per-agent page with observability, errors, tools, versions, run search, and the agent's full setup.

• A Health view with a heat map and failures by cause, plus a Releases timeline.

6. Evidence-Driven Agent Quality

Prove an agent is better before it goes live.

Voice agent experiments. Compare voice agents the way you compare text agents, against a simulated caller.

• Four comparison types. Agent Version, Model, Prompt, and Caller conditions (background room, noise level, and phone line).

• Conclusions you can trust. Each comparison leads with what changed and whether results got Better, Worse, Mixed, or stayed the Same. Better or Worse is called only after a paired permutation test; otherwise the result reads No clear difference. An AI analysis summarizes the results.

• Replay to the failure. Call replay with timed turns, tool calls, the checks decided at each turn, and Jump to failure.

• Voice-specific metrics. Response rate, interruption rate, caller audio condition, reply latency (p50 and p95), and cost per case.

• From result to fix. Suggest a fix for failed cases, compare pass rates across repeats, save a winning prompt as a new version, and export the comparison as HTML or PDF.

Task Map (beta). See the work your agents actually do, then benchmark it.

• Real work, mapped. Analyzes an organization's real runs into workflows, task clusters, and outliers, using a read-only, masked, organization-bound toolset.

• From map to benchmark. Turns the analysis into a benchmark proposal you can redraft, plus a printable report.

• Transparent cost. Analysis spend is shown in OU and billed as it is incurred.

Evaluation and benchmarks.

• Server-side Evaluate. Grades come from the stored run, and the judge sees the full input, real user turns, and attachments.

• Auto-run evaluation. Grade every completed turn automatically (off by default).

• Not graded, not zero. A judge failure shows as Not graded, never as a zero.

• Built-in benchmarks. Benchmarks are provisioned for each organization, each run is checked against its budget before it starts, and uploaded benchmarks are treated as untrusted.

More New Capabilities

Capabilities Hub

Skills and connectors in one hub under Agent Studio, with the right action for each role.

• One primary action per role. Members add to My skills, developers choose Use in an agent, and admins switch a skill on for everyone. Each skill shows how many agents use it.

• Clear ownership of connections. Members connect their own accounts and request connectors. Org-wide setup lives in Administration › Connectors, which shows the impact before a connector is switched off.

• Visible state. Each agent shows whose account it acts as. Expired, switched-off, needs-setup, and provider-rejected connectors show that state.

• Declared needs. Skills can declare the connections they need. The Plugins tab is a read-only preview of upcoming plugins.

ACTAVA-Built Agents

An agent ACTAVA builds arrives with its overview, first checks, and approval gates.

• An overview from the start. ACTAVA saves a summary, sample questions, and use cases once the draft has its key details.

• Knowledge in the right place. Files attached in a build conversation go to the agent's knowledge base, and ACTAVA asks before searching organization-wide.

• Starter checks. Checks run against the draft with editable expectations, and ACTAVA proposes value drivers and KPIs after a yes.

• Approval groups by role. ACTAVA can create an empty, named approval group for a step; only an org admin adds approvers.

Pipeline Builder Expansion (beta)

• AI pipeline wizard. Describe a pipeline in words. The wizard asks clarifying questions first, can revise from a finished run, and shows a before-and-after diff.

• New step types. Transform, External API, Run pipeline, Stop, Note, Read file, Model Selector, Synthesize data, Skill, MCP tool, Agent AI, Voice Agent AI (simulate a call in text or real audio), and email and notification actions.

• Operational controls. Write-only secrets per pipeline, disable a step, stop a run immediately, drag to reorder, export and import, incoming webhook triggers, and templates.

• API. /v1/pipelines and /v1/pipeline-runs are available in every environment.

Partner API and KORA MCP

More of the dashboard, from an API key, always within the key owner's role.

• Act as members. A WRITE or ADMIN key can be granted delegate (Can act as members, off by default) to run agents with act_as.

• Identity and format. GET /v1/whoami reports a key's identity, and GET /v1/agents/card-format returns the agent format the server enforces.

• Evaluate. POST /v1/sessions/{id}/evaluate grades a stored run server-side and reports its cost.

• Agent Workspace. Author a workspace, export one that rebuilds in another organization, set run conditions and early stop, and judge each step against its own rubric.

• Knowledge, skills, and connectors. Manage knowledge base files, upload skill versions, and set Microsoft 365 and Google Workspace switches.

• KORA MCP. Decisions, approval groups, schedules, and decision events through the key, with each decision recorded as the key creator's own vote.

Voice and Telephony

• Per-organization phone numbers. Each organization registers its own numbers. Inbound calls route by the number dialed, with no fallback to another organization.

• Voice Agent Simulator. Simulate a call from Voice Studio with selectable voices, consented voice cloning, noise rooms, and a back-and-forth script.

• Text to Speech and ElevenLabs. A TextToSpeech tool writes MP3 audio to the workspace, and organizations can bring their own ElevenLabs key and use ListVoices, SpeechToText, and SoundEffects.

Microsoft 365

• Personal by design. Every Microsoft 365 tool uses the account of the person the run executes as, never a shared credential.

• Per-service switches. Outlook Mail and Calendar, OneDrive, SharePoint, Contacts, Excel, Word, Teams, and Meetings each have an org-admin switch.

• Meeting transcripts. Agents can retrieve Microsoft meeting transcripts once an admin opts in.

Complete Feature List

Security and Access

• Pre-penetration-test hardening. No tenant-launched stdio MCP servers; backend MCP connections guarded against DNS rebinding; direct MCP tool calls limited to developers and above; background shells scoped to their session; no exception text in error responses; only system admins can change global ACTAVA knowledge.

• Role-scoped API keys. A key acts within its owner's current role, and a key whose owner is no longer an active member is refused (KEY_OWNER_CANNOT_ACT).

• Authorization fixes. Admin roles checked in the target organization; run lists and streams scoped to the caller; remote-execution toggles limited to org admins or the owning developer; invite routes no longer return tickets; SVG files download instead of rendering; hidden agents stay hidden.

• Signed identity. The router signs the identity it vouches for, drops stale elevated roles for removed members, and throttles repeated authentication failures.

• Sandbox floor. Agent Bash runs inside a sandbox floor that configuration cannot remove, without root capabilities and with an allowlisted environment, and file tools never follow a link out of the workspace.

• Sealed secrets. MCP bearer tokens are sealed in stored configuration and session snapshots, and client-supplied identity headers are stripped.

• Runtime-only run records. Only the agent runtime can save run records, and a save for another organization's run is refused.

• Honest attestation. The compliance attestation states only what the platform enforces, backed by tests.

• Identity hygiene. Stored user references use ACTAVA's own user id instead of the sign-in provider's, with a crosswalk for retired ids.

Observability and Audit

• Run integrity. A nightly keyed digest of each organization's runs detects changed or missing records, and runs and conversations are soft-deleted.

• Run provenance. Each turn records the code version, card, configuration received, and launch path.

• Audit export. A downloadable audit export from Compliance.

• Live-switch audit. Every change to an agent's live version is recorded.

• Truthful run events. Events report Bash failures, cache rate, and created files as they happened, and unresolved delivery warnings stay on the run.

• Judge alerts. A turn blocked by a judge gate can alert an org admin (opt-in).

• Last active. Member lists show when each member last used the app.

Agent Studio

• Card format enforced in code. The agent format is served from the code that enforces it, with knowledge scopes validated.

• No duplicate versions. Re-importing an unchanged agent writes no new version, and promoted versions carry their author.

• Portable approvals. Exports carry approval-group settings and connector identity, and older formats still import.

• Edit approval gates. Edit a tool approval gate in Studio, with one Remove approval control.

• Model picker. A compact picker with thinking effort; retired models are no longer accepted as overrides.

• Permission-aware wizards. Voice and Agent Workspace wizards follow the server's create decision and show errors instead of failing silently.

• Workspace fidelity. Per-step rubrics, run conditions and early stop, and completion refused while step work is unfinished.

• MCP Store redesign. An overview, installed servers in table or grid, catalog tiles, and automatic health checks.

• Hybrid knowledge search. Search finds exact identifiers as well as meanings.

• Knowledge readiness. A file is ready only once its content is searchable.

• Synthetic twin adapters. Agents can call a synthetic twin of an API with generated data, and REST adapters support unauthenticated APIs.

• Skill rules. One display-name rule on every skill write, Skill Builder for developers and above, and a file viewer with tree, formatted JSON, and line numbers.

• Value drivers. Adding a driver never replaces another.

Runtime and Chat

• Reliable resumes. Stopped turns recover, answered questions resume on the original model, and the page stays live after a gate or hand-off.

• Accurate run accounting. Tokens count across pauses, paused runs record only active time, and runs are saved even when setup fails.

• No stalls. A loop that only updates its to-do list ends the turn, and recursion limits are handled cleanly.

• More file types. Non-UTF-8 text such as Windows CSV, X12 EDI (.835, .edi, .x12), ZIP bundles, video, audio, and archives.

• Whole-page WebFetch. WebFetch returns the whole page, and crawling requires an explicit depth (1 to 5).

• Google Slides tools. Replace text and duplicate slides.

• Turn gate refinements. final_text_only keeps a checked turn's final text, and quoted figures count as cited.

• Inline media. Generated audio and video play in chat, and sandbox: links open the files an agent created.

• Resizable composer. Drag to resize, model choices survive a reload, and ?prompt= links stage text without sending.

• Clear environment. Agents are told where they can write and that Bash has no general internet access.

Voice

• Tools wait for answers. Voice tools wait for the caller's answer before running, and no tool runs on an empty required value.

• Clean hang-ups. The agent's last words finish playing before a hang-up, and the agent cannot hang up on its own unanswered question.

• Balanced recordings. Call recordings are a centered stereo mix of both voices.

• Privacy in voice. Sensitive data is redacted from voice logs, and room names and SIP identities carry no phone numbers.

• Rollout-safe calls. Live calls are protected during deployments.

• Large MCP toolsets. Voice agents keep MCP tools' optional parameters, and the tool allowlist handles more than 200 tools.

• Voice latency in the SLA. Per-turn response latency is recorded and reported at the 95th percentile.

Evaluations

• Updated judges. Heavy runs on Opus 5.5 with transcript caching, Regular on Sonnet 5.5, and Lite on Haiku 5.5.

• Metric refinements. A domain-neutral Safety rubric, Faithfulness that accepts input and tool results as sources, and per-metric effective settings.

• Comparison charts that add up. Pass rates count only the calls a check applies to.

• Benchmark management. Rename, re-describe, and re-tag benchmarks, and Simulation Environments with a shared blueprint canvas.

Platform and Reliability

• Database resilience. Connections recover after password rotation, every connection names its service, and a database outage is never reported as an invalid key.

• Stuck-run cleanup. One reaper ends stuck runs while sparing live turns, and runs with no progress end and are recorded.

• Long-run protection. Recovery uses liveness, not total session age.

• Capacity floor. Operators can set a timed capacity floor that batch sizing respects.

• Oversized tool results. Oversized MCP results are trimmed and retried instead of failing the turn.

• SLA measurement. Standard SLA availability with a monthly report and live status.

• Status page automation. Status is driven by direct checks of the Web App and Voice.

• Release integrity. Promotion requires benchmark runtime readiness, and rollback refuses an incomplete manifest.

• Dependency security. Critical and high-severity advisories patched across services.

Administration and Experience

• Grouped navigation. The grouped, job-named sidebar is now the only admin navigation.

• Security & Audit. A new sidebar group, and Settings › Security becomes Sign-in & devices.

• Focused developer rail. Voice Agent, Agent Workspace, Video Agent, and Run Governance are hidden from the citizen-developer rail; direct links still work.

• ACTAVA Settings console. An indexed settings console.

• Mobile layout. Navigation and panes work at phone and tablet widths.

• API examples. MCP Examples (beta), Pipeline Builder, skill, Decision Layer, benchmark, and webhook examples, plus Codex MCP setup.

• Webhook editor. Offers session.cancelled and schedule.fired.

• Accessibility. Error states instead of empty lists, accessible switch names, and AA-contrast success text.

Healthcare-Specific

• PHI encryption key rotation. Checkpoint encryption keys rotate with a rollback lever.

• PHI access records. Opening a session records a PHI read.

• PHI-safe logs. Database errors, OAuth failures, filenames, paths, and phone numbers are kept out of logs.

• HIPAA-mode test runs. Scripted voice test runs keep their transcripts for HIPAA-mode agents according to the agent's setting.

• Clinical benchmarks. MedXpertQA Text, HealthBench Professional and Hard, and CHI-Bench are provisioned for each organization.

• Claims files. X12 835 remittance files are accepted as chat attachments.

Deprecations and Removals

• Retired models. Kimi K2.6 retires to GPT-6.1 Sol at low effort, GPT-6 Sol to GPT-6.1 Sol, Sonnet 5 to Sonnet 5.5, and Haiku 4.5 to Haiku 5.5 (October 21, now hidden). GPT-5.6 Sol is no longer a default authoring option. Successors apply only where an admin has approved automatic upgrades.

• Navigation opt-out removed. The nav_v2 toggle and the flat sidebar are retired.

• Connectors move. Connectors leave the main rail for the Capabilities hub.

• Tenant stdio MCP servers removed. Only platform catalog entries may launch a stdio server.

• Shared Microsoft 365 credentials removed. Personal connections only.

• Cross-organization inbound call fallback removed. Calls route only to the organization that owns the number.

• Legacy Slack approvals deprecated in favor of the Decision Layer.

• Redirects. Agent Schedules pages redirect to the Scheduled hub where it is enabled, and /api-flows redirects to /pipelines.

• BATCH_TTL_HOURS removed. Batch deadlines derive from the work.

Read before you upgrade

• Review model consent. Automatic fallback is off by default and automatic upgrades wait for approval. Review both switches in model access.

• Model catalog. Regular is now GPT-6.1 Sol, and GPT-6 Luna bills at half of Haiku 5.5. GPT-6.1 Sol rejects reasoning none or minimal and any temperature, and Haiku 5.5 supports adaptive thinking only, so API callers that pin these models should check their request parameters.

• API keys follow their owner's role. Keys owned by members, developers, or people who have left may lose access to org-wide routes. Grant delegate only to keys that should act as members.

• Microsoft 365 is personal-only. Ask people who run Microsoft 365 agents to connect their own accounts, and review the per-service switches.

• Enable betas per organization. The Scheduled hub (schedules_hub), Decision Layer (decision_layer), and Pipeline Builder (api_flow_builder) each need a flag. Workspace review gates are enabled separately; contact your ACTAVA team.

• Controls rolling out per environment. Judge alerts, the allowlist filesystem view, and runtime-only run saves are enabled per environment after review.

• Plan who holds the brake. Only org admins and super admins can engage or release an Emergency Brake.

• Phone numbers. Per-organization numbers are provisioned by ACTAVA; confirm every inbound number is registered to the right organization.

• Batch sizing. The sizing baseline is 2,100 items per hour per organization.

• Stdio MCP servers. Move any tenant stdio MCP server to a hosted MCP server.

• Previews stay labeled. Auto-run evaluations are off by default, the new Agent Dashboard is a preview, and Plugins is a read-only preview.

Our commitment

Every month, ACTAVA ships a full release built around what our customers tell us they need. V9 alone brought more than 2,300 merged changes, six headline capabilities, and 73 further improvements, all built with the same engineering discipline: recorded design decisions, automated tests on every change, verification in staging, and security review.

We build that way because trust is the product. Agents that run unattended on real work have to be safe to stop, honest about what they did, verifiable by someone other than the model, and governed by the people accountable for them. V9 puts those guarantees into the platform itself, and we will keep raising the bar every month.

Thank you to every customer and partner whose feedback shaped this release.

Book a demo to see V9 in action with your own workflows.

ACTAVA.ai, The Healthcare AI Factory