How people assign work to agents, connect tools, set action boundaries, supervise activity and evaluate results.
This research area complements business workflow examples. It captures agent setup and operating patterns even when a post does not describe a complete business process or a new job role. Each source account stays distinct from any proposed enterprise practice.
Practitioner setup advice · Limited published evidence
Configure a long-running AI teammate
Ken Reiners describes a Perplexity teammate setup: define one job, connect only needed apps, scope GitHub repositories, set which Slack messages can be sent without asking, then try one job for a week and read each approval request.
Original post read on November 14, 2026. The post reports setup advice and product capabilities; it does not name a business task, show an executed run or report an independently checked result.
Source account
The setup sequence Ken recommends
The order follows the numbered setup advice and closing instruction in the post.
Person
Define the job
Write the teammate’s job in one line.
↓ next
Person
Connect required apps
Connect only the apps needed for that job.
↓ next
Person
Scope repository access
Choose which GitHub repositories the teammate can see.
↓ next
Person
Set message permissions
State which Slack messages may be sent without asking.
↓ then
Person
Try one job for a week
Read each approval request during the trial.
The post also describes the teammate as able to continue after the tab closes, send progress updates, follow up when something changes and message in Slack. Those are author-reported product capabilities, not observed task results or independently checked controls.
Proposed enterprise practice
Test one bounded task with explicit checks
Designed, not executed. This is a proposed pattern for a team to adapt with its task, security and system owners.
Task owner
Name one bounded task
Agree the expected output and an independent reference for checking it.
↓ next
Administrator
Configure limited access
Allow only the approved apps, repositories and communication actions in a sandbox.
↓ next
Agent
Work on approved records
Prepare a reviewable result and pause at actions outside its permission boundary.
↓ next
Independent reviewer
Check access and output
Compare observed access and result with the approved scope and reference.
↓ then
Task owner
Accept or hold the trial
Review the evidence for this task and decide whether the bounded trial met its criteria.
Failure or rejection: hold the result.
Repair the task or configuration, rerun the independent access and output checks, then return the evidence to the task owner. Stop or revoke access if the boundary cannot be confirmed.
The sandbox, independent review, permission verification, evidence retention and stop/revoke path are proposed controls. No enterprise environment, integration or acceptance test has been implemented or run.
Practitioner method · Limited published evidence
Give a writing agent context and feedback
Mike Fishbein describes an agent that uses a topic, a writing style profile and a person’s point of view to prepare a draft. The person edits it; published work and performance information then feed back into the writing context.
Original post read on November 15, 2026. Fishbein describes a method and claims that feedback improves drafts; the post does not provide independently verified performance results or name a customer deployment.
Source account
From topic to draft, with feedback
The sequence follows Fishbein’s numbered method. The feedback returns to the writing context for later drafts.
Agent
Receive a topic
Start from a topic idea.
↓ next
Agent
Gather writing context
Select a style profile and the writer’s point of view.
↓ next
Agent
Prepare a draft
Use the selected context to prepare a first draft.
↓ next
Writer
Edit the draft
The person edits the draft.
↓ next
Publisher not specified
Publish the content
The resulting content is published; the post does not identify who publishes it.
↓ then
Feedback
Update the writing context
Published content, edits and performance data feed back into the content brain for later drafts.
↩ Feedback returns to the writing context used for later drafts.
The post describes JSON style profiles in Supabase and a point-of-view markdown file. It does not specify organizational permissions or an independent approval process.
Proposed enterprise practice
Review a draft and its feedback
Designed, not executed. A fictional team would agree its own source material, permissions and acceptance criteria.
Content owner
Set a bounded brief
Choose a topic and approve the examples and source material available in a sandbox.
↓ next
Agent
Prepare a referenced draft
Use the approved material to prepare a draft and identify its supporting sources.
↓ next
Independent reviewer
Check the draft
Compare claims with source material and check rights, sensitive details and permitted access.
↓ then
Content owner
Accept or hold
Decide whether this exact draft is suitable for the agreed use; publishing remains a human action.
If the owner accepts and the content is published:
Content owner
Review feedback
Consider attributable performance and editing evidence before approving any change to the writing profile.
If a check fails or the owner rejects the draft, hold it.
Repair the draft, repeat independent checks and return it to the owner. Deny unauthorized publishing and hold profile changes until reviewed.
The sandbox, independent checks, owner decision and permission boundaries are proposed. They have not been implemented or tested in a business environment.