A managed teammate reviewed this site’s next explanation

Docs Quality, an existing AI teammate on our team, reviewed Socra’s product, pricing and setup pages and returned two briefs. We used them to correct how this site explains a first teammate’s cost and to add the Windows preview to the first-role guide.

Recorded October 17, 2026 in Seattle (October 18 UTC).

This is an executed assignment to our existing Docs Quality teammate, running Claude Code on its own Socra cloud computer. The receiver already existed, was connected and had access to the direct Channels conversation. No new teammate, subscription or access grant was created for this task.

What we used

Our Docs Quality teammate returned two source reviews. We used them to clarify the $20/month starting Socra bill and update our first-role guide to include the Windows preview. You can read the cost explanation and the first-role guide. The October 23 follow-up below records two further reviews and the corrections they informed.

Agents in the editor’s workflow supplied the questions, checked the findings, corrected source-attribution errors and published the changes.

Assignment

The editor asked the teammate through Channels to check public product, pricing and setup sources and return a release-note-editor readiness brief. The task specified the questions and a one-person, one-Agent, ten-hour cost example. It did not prescribe the answer or give the billing calculation. The agent’s existing role is checking public text against the product; this was a real review needed for the website’s new explanation.

Assignment delivered: 2026-10-18T03:19:28.786Z. Reply received: 2026-10-18T03:20:55.912Z, about 87 seconds later. These are message timestamps, not measured active work or billed duration. The activity record identifies this message as the run’s trigger and the Agent subsequently returned to sleeping.

What the first brief found

It listed what a teammate owns, its identity and access, how work reaches it, what the human still decides, and the cost of one person running one teammate for ten hours. It flagged that billed running time can exceed working time, that no page states the complete cost with the AI subscription, that pages differ on who can create a teammate, and that the quickstart's first task needed the Mac-only Channels app. The full text is in the appendix.

Correction returned by the same teammate

After the editor sent the attribution error back through Channels, Docs Quality rechecked both live pages and returned a correction at 2026-10-18T03:24:46.397Z: the product page gives an example; the model guide separately states the Codex default. It reported retaining the lesson in its own working notes, and explicitly reported no Cortex Module update. Those notes were not independently inspected. The later assignment below checks the returned work after that correction; it does not establish which memory mechanism produced the answer.

A second assignment for the same teammate

A fresh reader found onboarding unclear for people without a Mac. We sent the same Docs Quality teammate a new request to check prerequisites, documented operating-system paths, runtime choices and a useful first task. The request did not repeat the earlier example-versus-default correction.

Assignment delivered: 2026-10-18T03:43:32.505Z. Reply received: 2026-10-18T03:44:39.567Z, about 67 seconds later. These are message timestamps, not billed duration or human effort.

The brief identified a real mismatch: the Agent quickstart mentioned only Apple Silicon Mac, while its linked Channels install page also documented a Windows x64 preview. We independently checked both pages and updated the site's first-role guide to link to the current installer choices. The new brief also correctly cited the model guide for the Codex default. That is observable consistency with the correction, not proof that Cortex or a particular memory mechanism caused it. The task was sent in the same conversation.

What the second brief found

The quickstart was out of date: it named only Mac while its linked installer page documented a Windows preview. No Linux or phone path was documented. The Codex default model was stated and the Claude default was not. No page gave a total setup time or a one-page checklist. The full text is in the appendix.

What changed in the publication

The article explains $13 as usage drawn from a $20 seat’s included credit, rather than presenting $13 as a new customer’s entire bill. It names billed running hours, states the separate AI-provider cost, and makes the illustrative assumptions explicit. The first-role guide distinguishes defining a role from granting access or scheduling it.

The writing and website integration were performed by agents in the editor’s desktop session. Docs Quality supplied this source review; it did not write, deploy or independently verify the finished website. A separate fresh agent reads the live release afterward.

Evidence limits

These are two completed source-review assignments on an existing configured Agent. It does not measure hiring replacement, sustained throughput, cost savings, setup success, billable runtime, autonomous publication or durable learning across independent contexts. It supplies inspected work used in a real release. The release-note editor is a proposed role, not a claim that this Agent produced release notes.

Sources: https://workspace.socra.com/products/agents , https://workspace.socra.com/pricing , https://workspace.socra.com/docs/agents/get-started , https://workspace.socra.com/docs/admin/agents/billing .

Follow-up: two more source reviews, October 23

I wanted the handoff and cost explanations on this site checked against the current documentation. My coding agent gave Docs Quality two bounded assignments in separate Channels conversations. The teammate returned to idle between them. Its persistent model context remained available, so these were separate work sessions, not fresh-context tests.

The first review checked how another agent can continue work. It returned a sequence: connect each host, save decisions in Modules and unfinished work in Issues, then retrieve those records in the next conversation. The reviewed Grok guide covers Grok web and Grok Build; it does not document the Tesla experience I describe. The brief overstated two setup steps. The final text keeps administrator setup conditional and does not present saved instructions as mandatory.

The second review checked pricing and billing. It identified a missing distinction in the role page: Agents pause when credit runs out unless an Account owner enables additional usage. The monthly base charge continues until deletion and can exceed credit. We checked the source and corrected the role page and first-assignment guide.

The first reply arrived about 2 minutes 10 seconds after its request; the second, about 1 minute 17 seconds. Those are message turnaround times, not billed running time or a comparison of equivalent tasks. The second assignment being shorter does not establish improvement.

We could not attribute an actual running charge to these assignments from the available billing records. Human briefing and review time, and the coding agent’s coordination cost, were not measured. This record shows completed reviews and corrections used on the site; it does not establish cost savings or a rate of improvement.

Appendix: the returned briefs, as received

First brief

The text below is the Agent’s returned work. The cost and capability claims used in the article were checked against the linked public pages. An independent editorial review found one source-attribution error in the returned brief: the product page presents GPT-6 Sol as an example, not as a default. That statement is retained below as returned, with this correction; it is not used as a product claim on the article. The editor did not use the absence of a provider price as a product defect, or imply Admin permissions were tested. “Review every published note” is a proposed boundary for the example role, not a universal product restriction.

Readiness brief: delegating a release-note editor role to a managed Agent Checked 2026-10-18 (live Markdown, HTTP 200): workspace.socra.com/products/agents, /pricing, /docs/agents/get-started, plus the billing page linked from pricing (/docs/admin/agents/billing). Also CLI help for socra agent create and the docs page on choosing a model. Nothing was created, bought or changed.

What the Agent owns

  • Whatever responsibility you name when you create it. The pages do not describe a release-note editor role, so the human must write that job down (products/agents: "Give it a name and a responsibility").

  • Its work happens under its own identity. Source: products/agents "Start with a teammate and a responsibility".

Identity, computer, access

  • Own identity and its own cloud computer with a terminal; its home keeps files, instructions and memory across restarts. Source: products/agents "Its own computer in the cloud".

  • Access is only what you grant (knowledge, conversations, repositories). A connected AI subscription supplies AI usage, not resource access. Source: quickstart "Give it a place to work". A release-note editor will need read access to the change source you choose, granted separately.

How work reaches it

  • A Channels message, an assigned Cortex Issue, or a Task reminder. Source: products/agents "The work comes to your teammate".

What the human must still decide

  • The role and its limits; which knowledge, channels and repositories it can see; whether knowledge changes need your approval; which AI provider to connect (Codex via ChatGPT, or Claude Code) and its model and effort; whether to allow usage beyond included credit (off by default); and a review step for every published note.

Cost of one person running one Agent for 10 hours in a month (full month, 10 billed running hours)

  • Person: $20. That seat includes $20 of usage credit.

  • Agent: $3 + (10 x $1) = $13, drawn from the credit.

  • Out of pocket to Socra: $20; $7 of credit is left and expires at month end.

  • Plus your own AI subscription, which the pages do not price. Complete cost = $20 + that subscription.

  • Workspace has no trial.

Flags

  1. "10 hours" is ambiguous. Running time counts from when the computer starts until it goes back to sleep, rounded up to a whole minute per run. The pages do not say how long it idles before sleeping, so 10 hours of work can bill more than 10 hours.

  2. No page states the complete cost. Pricing says "Bring your own AI subscription" and the products page says plan allowance, but neither gives the provider price.

  3. Who can create an Agent differs by page. Quickstart says the Account Owner role is needed in Admin and owners and admins can use the CLI, while the products page "Create an agent" buttons say nothing about roles. I checked the docs, not the live Admin permissions.

  4. The quickstart's first task needs the Channels app, and its documented download is for Apple Silicon Mac only.

No numeric contradiction found: pricing, the billing page and the quickstart agree ($20 seat and credit, $3 per Agent, $1 per hour, the 2 Agents plus 14 hours example). The products page default model claim (GPT-6 Sol) matches the model docs (new Codex agents default to gpt-6-sol). Not verified: live Admin checkout, actual billing, sleep timing.

Second brief

Onboarding brief: starting a managed Agent when setup looks like a project Checked 2026-10-18 on the live public pages (Markdown, HTTP 200) and socra agent create --help. Nothing was created, bought or changed. Sources: Q = https://workspace.socra.com/docs/agents/get-started ; C = .../docs/admin/agents/connect-an-ai-subscription ; B = .../docs/admin/agents/billing ; P = https://workspace.socra.com/pricing ; I = .../docs/channels/get-started/first-message ; M = .../docs/agents/guides/choose-a-model ; D = .../docs/agents/guides/delegate-work ; K = .../docs/cortex/reference/cli

  1. Have before paying

  • A Workspace Account where you hold the Owner role to create the Agent in Admin; owners and admins can create through the CLI (Q). Connecting a subscription needs Owner or Admin (C).

  • An AI subscription with Codex access (ChatGPT) or Claude Code access (Claude), and the person who owns it present to sign in once (C). For ChatGPT, device-code sign-in must be enabled in ChatGPT security settings first (C).

  • Money: $20 per person per month (includes $20 usage credit) plus $3 per Agent per month plus $1 per running hour, drawn from that credit; no trial; extra usage is off by default; your AI subscription is billed separately by the provider (P, B).

  • For the quickstart's first task, the Channels app installed (Q, I).

  1. Windows, Linux, phone

  • Windows: documented. Channels has a Windows x64 preview, unsigned, so Windows may warn about an unknown publisher; Mac is the signed release (I). The CLI has a documented Windows install (K).

  • Linux: the CLI install is documented for macOS and Linux (K). No Linux Channels app is documented, and the quickstart's first task uses the Channels app, so a Linux reader has no documented end to end first-task path. Messaging through the Channels CLI is documented separately but is not presented as a quickstart alternative. Treat it as unverified for this purpose.

  • Phone: no phone or mobile setup path appears in the Agents, Admin Agents or Channels docs I searched. Missing answer; I did not test whether Admin works in a phone browser.

  • Contradiction: the quickstart says "The documented download is for Apple Silicon Mac" (Q), but the install page it links to also documents the Windows preview (I). The quickstart reads out of date.

  1. Model and runtime choices the docs establish

  • Two runtimes: Codex (ChatGPT subscription) and Claude Code (Claude subscription). The Agent uses the runtime that matches its connection (C). The CLI default is Codex (socra agent create --runtime codex|claude).

  • Model and reasoning effort are set per Agent in Admin or with socra agent update AGENT --model MODEL --reasoning-effort EFFORT; the choices come from the runtime and your provider account (M).

  • New Codex Agents default to gpt-6-sol (M). The default model for Claude Code Agents is not stated. Missing answer.

  • Agents on one connection share that subscription's usage limits (C).

  1. First useful assignment after the connection test

  • The connection test in the quickstart is one self-contained task in a direct Channels conversation: rewrite a given sentence and check it keeps the meaning (Q).

  • Documented next step: a bounded assignment with an observable result, with the responsibility, allowed material, what it may change alone, what needs approval and how to report (D). The docs example is checking a guide against the current setup flow and proposing corrections in an Issue, asking before changing pricing or access (D).

  • For a release-note editor, my suggestion (inference, not documented): give it the text of one recent change in the message and ask for a draft note, with no repository access yet, then review it before granting read access to a changelog source. Grant access only after the first result is reliable (Q, D).

  1. Missing answers that cause the "project" feeling

  • No page gives total setup time or a one page checklist; the quickstart sends the reader into a long connection procedure (C) before the first task.

  • Not stated: Claude default model, Linux Channels app, phone path, and how long an Agent stays running before it sleeps (affects cost).

Not verified: live Admin screens, billing, installers.