01 · Your own agents

Your team's agent, wherever you work, from day one.

Create an agent from what your team knows. Give it your repositories and the tools it needs over MCP. Connect a channel — portal chat, Slack, Discord, Teams, your Git host or your IDE over MCP — and start asking, or hand it a goal and let it work. Improve it later in chat.

Create your first agent See the setup
# platform-helpthread
P
Priya 10:02
@platform-agent why is checkout-api restarting in prod-eu-1?
platform-agent APP 10:02
The pods are OOM-killed: memory limit 512Mi, usage peaks near 600Mi since this morning's deploy. Our runbook says raise the limit and open an MR on deploy/checkout.
kubernetes · get_pods skill · oom-runbook
P
Priya 10:04
goal: raise it to 768Mi and watch for an hour
Plan written. Changing the limit asks first — approve it on the portal.
Setup

Four steps from nothing to an agent your team uses.

Create the agent

Name it, write the part of its main skill only your team can write, and add skills with their reference files. Or import an existing agent from Git.

Main skill — your part
You look after the platform team's clusters and deploy repositories. Prefer our runbooks over general advice…
oom-runbookdeploy-conventions+ add skill
Add its repositories

The portal reads each repository over SSH — fetch only — and drafts a reference and a skill. A person approves them before the agent sees a word.

deploy/checkoutapproved
infra/terraformready to review
platform/runbooksscanning
Give it tools over MCP

Connect any MCP server once, then tick the tools each agent may use. Each tool runs freely or asks a person first.

get_pods
kubernetes
runs freely
get_logs
kubernetes
runs freely
patch_deployment
kubernetes
asks first
search_issues
jira
runs freely
create_issue
jira
asks first
Connect a channel and start

Portal chat works at once. Slack, Discord and Teams each get one bot per agent, with a guided checklist. Your Git host gets comments on merge requests, and Claude Code or Cursor reach the agent over MCP. Then ask, or write goal: to hand it work.

✓Create the Slack app (or Discord or Teams bot) for the agent
✓Install it to your workspace
✓Invite it to #platform-help
Say hello in the channel
What an agent holds

It gets better the more your team uses it.

Every Claude Code session in your organisation can read the same agents over MCP, so know-how stops living in one person's head.

Skills

A main skill your maintainers own, plus focused skills with their files.

Repositories

Scanned into a reference and a skill, kept current as the code moves.

Memory

One log of what it learned, mirrored to your own memory store if you like.

Rules

Suggested from review verdicts and sessions; maintainers decide.

Ready-made kinds

Start from an agent that already knows the job.

Kubernetes agents

Many clusters per agent, each with a name and environment. Reads freely; every change asks first.

The reviewer

Seeded at signup. Reviews merge requests against PRDs — and you can teach it like any other agent.

Background job agents

Five seeded agents for the portal's own jobs — drafting, scanning, summaries — each one yours to tune.

Your first agent can answer your team today.

Create an agent