How It Works

The platform turns conversations into action through a five-step pipeline. You talk to an agent in Slack, Jira, or the Dashboard — it figures out what needs doing and gets it done.

For the full technical architecture, see the Architecture reference.

The pipeline

1. Trigger

A user starts work in one of four ways:

Slack: @buck, @dolly, or @pearl in a channel or thread
Jira: Assign a ticket to Buck, or comment on a ticket he’s working on
Dashboard: Create a session directly from the web dashboard
Pull Requests: Open a PR (automatic review) or comment @buck on a PR (Buck only)

2. Event routing

Events arrive as webhooks and are published to NATS JetStream subjects (webhooks.{agent}.{type}). Each agent has its own consumer that pulls events from its namespace. Per-session ordering ensures events for the same conversation execute sequentially via sharded mutexes.

3. Agent reasoning

Each agent uses Claude (via AWS Bedrock Converse API) to reason about the event:

Buck: Searches code via Sourcegraph, checks observability via Grafana, creates Jira tickets and dashboard sessions, dispatches Claude sessions to EKS pods
Dolly: Reads specs from Confluence and Google Drive, decomposes epics, creates/updates Jira tickets, generates status reports, manages handoffs to Buck and Pearl
Pearl: Reviews designs, generates Svelte components, integrates with Figma, publishes interactive snippets, manages design languages and component library

4. Execution

Buck: Dispatches an ephemeral EKS pod with Claude Code for implementation. Pods run in tmux with terminal access, MCP plugins (Sourcegraph, Grafana, PostHog, Playwright), and full codebase access. By default, Claude decides what to do — in plan mode, it only produces a plan for review.
Dolly: Operates inline — reasons and responds directly via Slack/Jira. Posts decomposition approval UIs with interactive Block Kit buttons.
Pearl: Operates inline — reasons and responds directly via Slack/Jira. Publishes interactive snippets to the dashboard for preview.

5. Callback / Response

Buck: The pod reports results back to the agent via runner API (pod-bound JWT). The agent posts the plan or PR link to Slack/Jira. Sessions can pause (saving state to S3) and resume later with follow-up instructions.
Dolly: Responds directly in the Slack thread or Jira comment with reports, decompositions, or status updates.
Pearl: Responds directly with design deliverables — reviews, generated components, UI copy, or snippet preview links.

Modes

Mode	Behavior	Applies to
Default	Claude reads the task, plans, and implements in one session	Buck
Plan	Claude reads the codebase and posts a detailed plan for human review	Buck

Interactive capabilities

Capability	Buck	Dolly	Pearl
Terminal access	Yes (WebSocket tunnel from dashboard)	—	—
Pause/resume	Yes (S3 state persistence)	—	—
Real-time status	Yes (SSE pod status updates)	—	—
Block Kit interactions	Yes (plan approval buttons)	Yes (decomposition approval)	—
Multimodal input	Yes (images from Slack threads)	—	Yes (design images from threads)