Hours-long agents need a control plane, not a longer request
Vocion Teamvocion-v2.47.9
ADR 0004 is accepted on main; phase 1 (worker_run, the worker API, the reaper) is in review as vocion-core#237, a draft PR, not merged.
How-to write-ups, release notes, and notes from the Vocion team.
ADR 0004 is accepted on main; phase 1 (worker_run, the worker API, the reaper) is in review as vocion-core#237, a draft PR, not merged.
Why a chat interface is a new trust boundary, and the narrow Slack phase-1 slice now in review as vocion-core#238, not merged.
Running Vocion's marketing entirely on agents for 24 hours: one machine, four models, human-in-the-loop, every change a PR.
Autonomy levels are usually an analogy. Here are three real missions at levels 1, 2, and 3, the code that enforces the gate, and what the levels do not do.
What an open-source agent workforce platform is, what it must include, and how Vocion implements it with git-authored agents, approvals, and audit.
What to trace when many agents run: per-org, per-agent, per-feature cost, tool calls linked to a workflow, and the schema behind it.
A build guide with the artifacts: the agent file, the team file, the mission, the workflow with a human gate, and the version row the apply writes.
A small team wants shared agents, shared memory, BYOK and an autonomy dial. Here is what assembling it costs, and what one install still lacks.
The exact commands a coding agent runs to install Vocion, scaffold a workspace, apply it, and start the server, with a check that proves each step worked.
A guide to deploying AI agents in production with human approval: five insertion points, an autonomy ladder, trust thresholds, and a REST review queue.
Build a support agent that summarizes and drafts ticket replies, then holds them in a review queue until a human approves.
Vocion's review queue API: GET /api/v1/reviews lists an org's pending agent actions, POST /api/v1/reviews/decide approves, rejects, or edits one.
An agent can extract and score an application, but only a person should reject one. Here is the object, the gate, and the API call that enforces it.
When a human edits, approves, or rejects a draft, that decision can become a rule the agent reads next time — but only after a person approves it.
Which version of the agent did that? A concrete schema for AI agent versioning and audit trails: version rows on apply, a workspace SHA on every tool call.
"AI agent registry" means two things: a discovery directory of running agents, and a versioned source of truth for agent definitions. Which one you need.
How to build an AI agent approval workflow you can self-host: an approve step, a skill that requires review, trust thresholds, and a REST decision API.
What self-hosting an agentic workflow platform actually runs: Postgres, Temporal, Langfuse, an app server. Compose files, ports, RAM, no SaaS trial.
v2.26.0 gives every lead a dedicated page at /gtm/lead/{hubspot_id} — the decidable card, the research record, and the decision trail on one addressable URL.
v2.24.0 closes the API gap — fourteen new /api/v1 endpoints give review, feedback and learnings full parity with the dashboard, and admins can now mint API tokens without shell access.
v2.23.0 adds a Strapi connector that discovers your collections from a URL and a token, ingests many collections through one source, and survives a collection it cannot read.
v2.25.0 lands review card contract v2 — subject, provenance, recommendation, typed content and per-object verbs — so any object type presents through one template and decides through one path.
v2.16.0 serves the agent domain-tool registry over MCP — typed CRM reads, cache-aware Zoom transcripts and Gmail threads, knowledge search, and review-gated writes — plus a CLI for issuing Bearer tokens.
A working setup for the v2.16 MCP surface in Claude Code: registration without leaking a token, exact CRM counts, on-demand call transcripts and mail threads, and writes that always wait for a human.
v2.15.0 gives agents three structured HubSpot read tools — one per object type — with exact COUNT(*) totals, discoverable facets, explicit pagination, and honest refusals when a field isn't synced.
Vocion now detects which of your recorded calls were discovery calls — without ever letting an agent read a transcript that didn't first match a CRM record you own. Plus: automations become inspectable, testable, and owned.
v2.3.0 ships the core base pack: versioned RevOps agents a workspace inherits with extends: core and overrides in YAML — with per-layer provenance and a guard that stops any override from loosening an approval gate.
The Jira connector lands — projects and issues as retrievable knowledge, hourly incremental — alongside per-source reconcile schedules: a second full-sync cadence, on every connector, that prunes what incremental syncs can never see deleted.
The review queue becomes a focused one-at-a-time flow with editable, steerable actions — and briefings go team-scoped, with a workspace rollup, history, and live regeneration.
vocion-v2.1.5 makes agent chat legible and safe: a typed hierarchical Activity trace with tool drill-in, true token streaming, inline [n] citations with a two-level Sources drawer, recommended-action cards that land in human review, and a never-auto-send guard on outbound email.
Teams become first-class: authored in YAML with a lead and an accountable human, rendered as an org chart at /dashboard/teams, consulted by the workspace lead in chat and missions, and seedable from a bundled sample workspace.
The first breaking release removes the bundled workspace from vocion-core. Workspaces now live beside the checkout as directories you own, scaffolded with one command and located via WORKSPACE_PATH.
Long-term memory lands: agents extract facts and preferences from past conversations and recall them in brand-new ones, with continuity that survives sessions.
The BYOA runtime ships the agent loop as a standalone artifact: one HTTP contract, hostable on a laptop in dev or AWS AgentCore Runtime in the cloud, with a CI deploy pipeline.
Four verbs, no overlap: talk to the team, decide what needs you, observe everything it did, find anything. Activity is new — one feed for every check, run, fire, and sync.
Missions are now charters a team owns continuously — checked on a schedule, discharged with only the work that's needed. Plus schedule triggers for workflows, agent steps, and an Automation page for every clock in your deployment.
Mission checks now remember what they flagged (threads age across checks), high-confidence proposals can execute on enabled trust rules, workflows can ask a human for input mid-run, and briefings publish to their own page.
Agents are who. Missions are goals. Workflows are procedures. Automations — new and first-class — are when. Plus: watch your agents actually think, with streamed chain-of-thought.
Restrict any source to specific members — enforced in chat and search as an intersection, never a leak. Plus e2e:demo — the full briefing + CRM sweep as an executable test, passing 9/9 against live production data.
Agents now sweep the last 48 hours of email, detect what moved a deal forward or backward, and queue scored CRM updates for one-click human approval — the full loop, running on live HubSpot + Gmail data.
Vocion deployments are multi-user from day one — a Members page to see everyone, manage roles, and mint one-time invite links that need zero email infrastructure.
Credential onboarding lands — paste a token or run one Google OAuth consent and your connectors sync for real. And agents gain propose_action: CRM updates and emails proposed with confidence scores, gated behind your review queue.
The dashboard brand is now a slot — set a wordmark, a subhead, and your own glyph via build-time env, so a Vocion deployment wears your identity and the OSS build carries no one else's art.
Pending work now routes to a specific person — per-person queues, an unassigned triage lane, and snooze — so a workspace with many humans can actually divide the approvals.
The second action on the framework — update a deal's stage, a contact's fields, a missing next-step — proposed by an agent, approved by a human, executed against the CRM.
The trigger runner. An inbound event — a prospect reply, a meeting ending, a deal changing stage — fans out to the workflows subscribed to it, deduped and audited. The last load-bearing piece of the loop.
Agents now carry a role, a work mode, and a team — so a deployment reads as an org chart: one Lead you brief, its specialists doing focused work. A new Teams view makes it visible.
The write half of the runtime. A narrow, typed action (send an email, update a record) that an agent proposes and a human approves — the same permission model as everything else, now applied to mutation.
The control plane gets its first write surface — the unified review queue, list and decide, on tenant Bearer tokens that resolve into the same permission model as a human.
A source's schedule now actually fires — runSync runs inside a durable Temporal Workflow on a Schedule, so connectors keep their knowledge fresh on their own cadence, crash-safe and incremental.
A few days ago we wrote about the platform direction FirstHQ was forcing. Here's the first wave, shipped: client/team-scoped retrieval with document ACLs, a discovery-vs-mutation permission model with one review queue, and durable, resumable, incremental ingestion.
The agent/tool plane goes multi-tenant. A single Streamable-HTTP MCP endpoint where the org is derived from a tenant Bearer token — every tool call scoped to that org, under the same permission model as a human.
We're building FirstHQ (an AI team HQ for agencies) on Vocion — and using it as a forcing function for the platform. Here's the architecture direction it's driving: scoped retrieval with document ACLs, durable ingestion pipelines, a discovery-vs-mutation permission model, and a clean split between the MCP tool plane and the API control plane.
The Drive connector — Docs, Sheets, Slides, and text files exported into scoped, retrievable knowledge — completes the connector set the two reference deployments run on.
Two more waves of platform work shipped: tenant API tokens that resolve into the permission model (the control plane), and a connector pack — Google Ads, GA4, HubSpot, Gmail, Slack — on the durable, incremental ingestion pipeline.
The credential vault is now wired into the ingestion pipeline — a connector's OAuth/API credentials are stored encrypted-at-rest per tenant and resolved at sync time, so the connector pack can actually authenticate.
vocion-v1.25.0 adds Missions: hand your AI team a goal in plain language and watch them plan it, split the work, produce artifacts, and pause for your approval. One Agent framework, three work modes. What works gets promoted into a Workflow.
v1.24.0 gives every agent a general toolbelt out of the box: live web search, page fetch + crawl, image generation, a safe calculator, and artifact creation. Providers are pluggable; defaults work with minimal setup.
Use SaaS for generic tasks. Use Vocion for the AI workflows you need to own, audit, and extend. Not anti-SaaS — anti-lock-in.
Vocion is open source under MPL-2.0, so your company can self-host it, modify it, embed it, and commercialize what you build on top. Open framework underneath; proprietary value on top.
Turn repeatable expertise into scalable, AI-enabled delivery — without removing human judgment. Which workflows are good candidates, and how the path works.
The fastest path to AI value isn't a brand-new process — it's improving a workflow your team already runs every week. A six-step way to do it.
The best AI workflows don't just save a few minutes — they change the operating model. Where AI creates leverage, and why the workflow matters more than the model.
A walkthrough of the v0.4.0 support-reply demo (two operations, one workflow, one card) — historical shape; see the staleness note before copying any YAML.
vocion-v0.3.0 ships ChatShell — a server-sent-events chat surface with 15s keepalives, thinking summaries, and inline tool calls. Here's how to use it and what's behind it.
The open-source Agent Workforce Platform. AI coworkers you can trust with real work — running in your own AWS account, on a foundation your company owns.
MPL-2.0 · Self-hosted · Runs on AWS