Coworker

About thirty five connectors are named across two pages, against a stated "50+ OAuth connectors" and a directory at /connectors.

On the platform page they cover CRM (Salesforce, HubSpot), communications and meetings (Slack, Gmail, Zoom, Google Meet, Google Calendar, Microsoft 365), docs and project work (Google Drive, Google Sheets, Notion, Confluence, Jira, Linear, Asana, Figma), engineering (GitHub, Datadog, PagerDuty, VS Code), data (Snowflake, BigQuery, Amplitude), payments (Stripe), the browser (Chrome) and the systems of record for IT service management, HR and finance (ServiceNow, Workday, ADP, SAP, NetSuite).

The agents page adds Gong for sales, Zendesk and Intercom for support, and Microsoft Teams. SAP, Workday, NetSuite and ServiceNow are not trivial connectors to build.

Data moves both ways, and the vendor says so plainly: "Every connector is read and write. Agents pull data and push updates across CRM, comms, support, docs, code, and data." The agents page names the writes ("Updates CRM. Creates tickets. Sends follow-ups. Drafts reports."), and the published run drafts into Gmail, updates Salesforce and posts to Slack. Changing a deal record in someone else's CRM changes state in a system Coworker does not own.

Agents "inherit the permissions already in place," with no new logins and no new SSO maps, so there is no parallel access model to drift, which is what makes fifty read and write connectors safe to grant. Chrome and VS Code appear in the OAuth connector list as ingestion paths feeding the graph, not as interface control.

Sourcecoworker.ai/platform integrations section and /agents connector listread 2026-09-12

The Agent Builder is the authoring surface, and the customer uses it directly: "Create custom agents in minutes. Define what they do, which tools they use, and when they run," with no code required, and triggers, approval gates and notification preferences set through a visual interface. What the agent does, which tools it uses, when it runs and where a person steps in are the parts of a workflow, and the buyer sets them all.

The sequences span several steps and systems, and they branch. The account agent's published run is three ordered actions across three systems with a condition in the middle: it "drafted outreach for six accounts" through Salesforce and Gmail, "escalated one strategic account to AE" through Slack, and "posted expansion summary to Slack."

One of six accounts went to a person, a branch on the state of the case. The Incident Orchestrator "routes incidents, enforces escalation timers, tracks resolution": routing is a decision, a timer is state held across a gap the agent does not control, and tracking to resolution is a long running commitment.

State persists across runs by design: agents run 24/7 in the background without anyone stepping in, and the platform describes orchestrating several agents from a single control plane. The participants are distinct: the agent and the approver at a gate, the AE an account escalates to, and the connected systems each step reads and writes.

The claim of end to end execution is specific: "Other AI tells you what to do. This one does it. Updates CRM. Creates tickets. Sends follow-ups. Drafts reports." No page describes retry, failure handling or what an agent does when a step fails.

Sourcecoworker.ai/agents capabilities, agent examples and FAQ, with /platformread 2026-09-12

OM2 is maintained by design. It "reads every connected tool continuously," with no uploads or re-syncing, and "learns backwards over historical data at setup, while continuously evolving as it sees more each day." The vendor draws the contrast itself: most assistants, it says, "fetch document chunks at query time and forget afterward."

The retrieval design is described piece by piece. Knowledge is stored as connected facts about people, projects, customers and decisions, with customers, projects and deals, as well as people, as first class nodes rather than raw text. Recall is synthesized ahead of time rather than found at query time, "relevancy-ranked by proximity, recency, and centrality," with "sub-second graph traversal." Retrieval is scoped by "permission-aware traversal, built into the graph," with access policies on every fact and connection.

Answers are attributed: "every answer traces back to its source," and agents get "the right permissioned context across Slack, CRM, docs, Jira, and 50+ apps." The agents page lists the corpus as messages, tickets, deals, docs, sheets, calendars, CRM, emails, meetings, transcripts, code, wikis, boards, PRs, recordings and contacts. The vendor's comparative figures (84% preference, 20% faster, 66% cheaper) are its own unaudited benchmarks.

Sourcecoworker.ai/platform OM2 section with /blog/introducing-om2 and /agentsread 2026-09-12

Approval gates are a named product capability that customers configure: "Add approval steps to any agent workflow. Review, edit, or reject actions before they execute." A reviewer can amend the agent's proposed action before it fires, not only approve or deny it, and the hold sits on the agent's own action.

Both positions on autonomy are stated: "Both modes are supported. Agents can run fully autonomously for routine tasks, or you can add approval gates to any workflow so a human reviews and approves actions before they execute." Background runs work the same way: "get notified when they're done, review the output, and approve actions before they go live."

Four controls are named together: permissions and approvals, credit limits and escalation logic. A credit limit is a hard spending stop on an autonomous agent that does not depend on anyone noticing. Escalation logic shows up in the products too: the Incident Orchestrator "enforces escalation timers," and the account agent "escalated one strategic account to AE."

Agents "inherit the permissions already in place," with no new logins and no new SSO maps, so an agent cannot go beyond the authority of the person it acts for.

Sourcecoworker.ai/agents capabilities and FAQ, with /platformread 2026-09-12

Attestation and access control are both published. Coworker holds SOC 2 Type II, GDPR compliance and CASA Tier 2, shown as three badges on the platform page that each link to a Secureframe trust portal at village-labs.secureframetrust.com. Under Identity and access the page states SSO, RBAC, and encryption at rest and in transit, and the pricing page adds SSO to the Enterprise entitlements.

Permissions reach down to each fact, deeper than the usual RBAC line. Coworker is "permission-aware by construction," with access policies on every fact and connection, so everyone sees only what they are allowed to see, through traversal of the graph that respects permissions. Agents "inherit the permissions already in place," with no new logins and no new SSO maps, so there is no privilege elevation and no parallel access model to drift. For a product that pulls Slack, CRM, email and HR systems into one graph, authorization on each fact decides whether the whole thing is safe.

Isolated single tenant infrastructure and complete audit trails across the platform back this up, along with a stated commitment never to train models on customer data and published legal agreements, namely a privacy policy, terms and a Master Services Agreement.

The agents page claims agents "tap into 100k+ deployed agents' learnings," which reads as learning across customers, against the flat statement that customer data is never used to train models. Both are the vendor's own words, and nothing published says whether the shared learnings are behavioral patterns rather than customer content.

Sourcecoworker.ai/platform enterprise-grade section and FAQ, with /pricingread 2026-09-12

An action log is named. Under "Your rules, full visibility," Coworker promises guardrails you set, "audit trails you own" and "every action logged, every decision auditable." The agents FAQ repeats it for autonomous runs, "every action is logged for full auditability," and the platform page lists "complete audit trails across the platform" among its enterprise controls beside permissions and identity. Logging every action records what the agent did, not only what it produced.

Answers are attributed separately: "every answer traces back to its source" in the OM2 layer, so a conclusion can be walked back to the fact and the connected tool it came from.

The audit trail is asserted rather than described. No page states its retention period or whether it can be exported or filtered, and whether it captures reasoning as well as actions, or exists below the Enterprise tier, is not stated either. For agents that write to CRM and HR systems under credit limits, it is the one enterprise claim here with no detail behind it.

Sourcecoworker.ai/agents and /platform enterprise-grade section, with /blog/introducing-om2read 2026-09-12

The graph is ingested, so it is knowledge rather than memory the agent writes. OM2, which replaced OM1 as the headline layer in early September 2026, is built by reading the customer's connected tools continuously, with no uploads or re-syncing, and it learns backwards over historical data at setup.

Parts of it do behave like memory. Agents are described using the graph "to track progress, learn from corrections, and continuously improve," and the third how it works step reads "memory compounds with every interaction"; learning from corrections is something the agent writes, though it is also learning absorbed into the product. Scoping is strong: isolated single tenant infrastructure, access policies on every fact and connection, and permission aware traversal, with permissions per user in the native apps.

Two things are missing. No retention period or expiry is stated for the memory, and no purge or deletion path; the only retention figure published is 90 days for meetings on the Enterprise tier, which covers recordings rather than the graph.

And Coworker MCP is "one connection, read-only by default": the graph can be read, but nothing documented lets a customer or an outside caller edit, correct or delete a fact in it. The store is scoped, can be inspected through the MCP read surface and the native apps, and carries permissions on each fact. For a vendor whose product is organizational memory, the missing retention and correction controls stand out.

Sourcecoworker.ai/platform, /agents and /blog/introducing-om2read 2026-09-12

Four ways to deliver it are named in one line on the platform page, all on isolated single tenant infrastructure: cloud, private cloud, on premises or air gapped. On premises and air gapped installs sit in the customer's own environment, so the customer decides where data rests. The FAQ says it separately: "SOC 2 Type II, GDPR, and CASA Tier 2 verified, with SSO, RBAC, encryption, audit trails, and air-gapped deployment support."

Single tenancy is the baseline, not an upgrade: the OM2 announcement says the memory "runs on an isolated single-tenant infrastructure in our cloud or via VPC." VPC peering and bring your own model on the customer's own infrastructure are further options, so a buyer who cannot send data to a shared service has three published routes.

The homepage says routing is "all hosted in the US," and the OM2 post leads with cloud and VPC, which reads as hosted only; the full range is stated on the platform page.

No named region list for the hosted option is published, and no residency selection for cloud customers, nor any page on what an air gapped deployment costs, excludes or requires. Routing across twelve providers is hard to square with an air gapped install, and nothing addresses that; bringing your own model on the customer's infrastructure may be the answer, though it is not stated.

Sourcecoworker.ai/platform OM2 security line and FAQread 2026-09-12

Coworker publishes a named catalog, with each agent described and tagged by role. Five agents carry a function, a job and a stated outcome. Expansion Pipeline, for Account Management, "finds accounts hitting usage ceilings and generates upgrade outreach," and Revenue Recovery, for Finance, "processes payment failures with segment-specific recovery playbooks." Pipeline Cleanup, for Sales Ops, enforces CRM discipline on stale deals, past due dates and missing data.

Incident Orchestrator, for Engineering, "routes incidents, enforces escalation timers, tracks resolution," and Data Analyst, for BI and Analytics, "queries your data, builds reports, and surfaces insights automatically." Three more are named in the deployment section, Post-Meeting Actions for Sales, Contract Review for Legal and Bug Triage for Engineering, and the agent picker offers Account, MCP, Engineering, BI, Legal, Meetings, Finance and Sales as classes to choose from.

Each stands alone: remove Revenue Recovery and Incident Orchestrator is untouched, remove Bug Triage and Pipeline Cleanup still works. They serve different functions, on different objects, for different buyers in the same company. The product learns in minutes once a tool is connected and suggests agents suited to the user's role, so the catalog proposes its own members from what a customer connects.

OM2 carries shared skills and artifact templates that travel across models and do work when selected, and use case pages by function make the catalog browsable, one each for Sales, Engineering, Customer Success, Support, Product, Marketing, People, Operations and Industries. The named agents are cards rather than individual pages, and no count of the full library is given.

Sourcecoworker.ai/agents agent set and role suggestions, with /platform OM2 skillsread 2026-09-12

Schedules, events and manual starts are all listed among the capabilities. Scheduled triggers ("set it and forget it") run agents for daily briefings, weekly reports and hourly monitoring. Custom triggers ("react to events instantly") start agents on new Slack messages and Jira updates, CRM changes and calendar events, four event sources in four systems, firing on the customer's data rather than the clock. The FAQ states the three routes: on a schedule (daily, weekly, hourly), on real time events, or manually.

Agents work where the team already works. Coworker "lives where you work" rather than in another tab, running in Slack, joining meetings and updating boards. The Coworker Native Apps (Work, Chat, Build, Meetings, Agents, Artifacts) are first party surfaces, and Coworker MCP puts the same agents inside Claude, Cursor, ChatGPT and Windsurf. Email appears in the agent examples, which draft outreach through Gmail.

Agents run 24/7 in the background without anyone stepping in and send a notice when done, and all of this is generally available rather than held to one tier. No page says whether a customer can define a compound or conditional trigger beyond the four named event types.

Sourcecoworker.ai/agents capabilities and FAQread 2026-09-12

Three levels are documented, and the last puts the customer in control. Coworker "routes each task to the optimal model across closed, open-source, and self-hosted providers," balancing cost, latency and accuracy, with "every task routed to the cheapest model that meets the quality bar." It names "12+ model providers," among them OpenAI, Anthropic and Google, plus open source Llama, Mistral and MiniMax. And BYOM lets a customer "bring your own models on your own infrastructure," so a customer can supply and run its own model.

The pitch is no lock in: "As better models ship, your team gets them automatically, no migration, no vendor lock-in," and OM2's memory and skills travel with the customer whatever model it chooses. The FAQ repeats the provider list. Coworker MCP, on the same page, is a separate fact: it lets the customer's own assistant reach Coworker's memory.

No page says whether an administrator can pin a specific model per agent or can only rely on automatic routing.

Sourcecoworker.ai/platform routing section and FAQread 2026-09-12

Coworker MCP is a named product with its own page: "Org memory, 50+ tools, and dedicated agents inside any MCP client. One connection, read-only by default, same SSO and permissions," with Claude, Cursor, ChatGPT and Windsurf named as clients. The OM2 announcement adds a second route, "OM2 connects via MCP or API," and the FAQ confirms it: "Can I use Coworker's memory outside Coworker? Yes, via Coworker MCP in any MCP-compatible client." That is the customer's assistant calling Coworker.

Dedicated agents inside any MCP client go beyond a read only knowledge surface: an outside assistant can invoke a Coworker agent, and those agents update CRM records, create tickets and send follow ups. The agent picker lists MCP as one of its named surfaces alongside Account, Engineering, BI, Legal, Meetings, Finance and Sales.

The connection reads by default, and the wording implies write access can be configured, but no page says how or what a write scope covers. Permissions and SSO come from the same identity, so the surface does not widen a customer's exposure. The memory is sold as something that leaves, "one memory, every surface," with skills traveling too. No tool list or endpoint reference is published, and no SDK or OpenAPI document, and the API named in the OM2 post has no published reference.

Sourcecoworker.ai/platform portability section and FAQ, with /blog/introducing-om2read 2026-09-12

Nothing puts a change under test. No sandbox, dry run, simulation or staging workspace appears, nor a holdout, A/B comparison or versioning and rollback of an agent configuration, nor an evaluator, scored result or regression surface, across the platform page, the agents page, the pricing tiers and the FAQ. The agents FAQ walks through the deployment lifecycle in detail, where a test step would appear if one existed: a customer builds an agent in the Agent Builder and it runs.

The platform page is organized around five layers, OM2, Routing and Portability along with Integrations and Enterprise Grade, and evaluation appears in none of them. The third how it works step, Optimize and scale, says "memory compounds with every interaction," which is the memory layer growing, not a change being tested.

Agents are said to "get smarter with use," to "observe your patterns and suggest new automations," and to "get smarter every week as they learn how your team works." An agent that improves is not an agent under test, and a suggestion is not a verdict. Approval gates let a person review, edit or reject actions before they execute, which checks each output, not a change.

Sourcecoworker.ai/agents and /platformread 2026-09-12

Every path runs through protocols. Data comes in through OAuth connectors feeding the graph, actions go out as connector writes that "updates CRM," "creates tickets" and "sends follow-ups," and the outside surface is an MCP server. The vendor frames OM2 as replacing work at the interface rather than doing it, with "sub-second graph traversal" in place of brute force tool calling. Nothing operates a screen, so a change to a UI element has nothing to break across fifty connectors that are OAuth APIs.

Chrome and VS Code appear among the OAuth connectors on the platform page, in a list introduced as connectors feeding the graph alongside Slack, Snowflake and Workday, and the agents page lists what is read from them, such as code, PRs and recordings. That is ingestion into the knowledge graph.

No hosted or local browser, desktop session or remote computer control appears across the site, and no RPA, recorder or extension that acts on a page.

Sourcecoworker.ai/platform integrations list and /agents context listread 2026-09-12