The automated team

ShipSignal has no human agent roster. Its “agents” are the LangGraph graphs and nodes that do the work — each a stage in the pipeline that turns a release into approved launch content. Each stage below describes its role, what it does, and the human approval gate it feeds.

Agent → capability mapping

Which artifact-type capabilities each agent is allowed to produce. The worker gates generation by this exact allowlist — an operator edit wins per agent, otherwise the seeded code default applies; a type removed here is never produced even if a run selects it. Removing every capability from an agent reverts it to the code default. Only content generation produces artifact types — the other stages emit pipeline outputs (evidence, scores, the demo video, skill candidates), so they have no editable capabilities. The artifact types themselves and the skills that ground them live in Capabilities.

Release Intelligence

Evidence → feature manifest

Graph: graph.py (release-intelligence graph)

• Collects release evidence from GitHub (diffs, PRs, issues, docs) and redacts it before persist.
• Runs deterministic signal extraction, then clusters and scores the signals into candidate features.
• Persists the feature manifest and halts for human review.

Feeds Gate #1 — the feature manifest is human-approved before any content is generated.

Review launches

Content Generation

Approved features → drafts + provenance

Graph: content_graph.py (content-generation graph)

• Loads the approved features and snapshots the active repo skills that will ground generation.
• Generates the selected artifacts in parallel (blog/changelog, one-pager, social, demo script, audio digest).
• Extracts each claim, links it to concrete evidence, then runs deterministic policy checks and Bedrock Guardrails.
• Persists the reviewable artifacts and halts for human review.

Feeds Gate #2 — artifacts are human-approved (a blocking check marks an artifact "blocked") before publish.

Review artifacts

Eval

Quality scoring of generated artifacts

Graph: eval_orchestration.py / eval_rubric.py

• Scores generated artifacts against the rubric with an LLM-as-judge, dimension by dimension.
• Applies regression gates so quality and cost/latency stay within budget across launches.

Advises Gate #2 — the scores and rubric breakdown inform the reviewer; they do not auto-approve.

See quality signals

Media Generation

Approved demo script → demo video

Graph: media_graph.py (media graph)

• Validates the click-path JSON, then drives a Playwright capture of the demo.
• Stores the raw recording, generates ElevenLabs narration from the approved script, and assembles the video with ffmpeg.
• Stores the finished media in S3 and records the media asset.

No human gate — it runs only on an already Gate #2-approved demo script; safety is structural (schema-validated click-path, materialized-audio guard).

Open a launch

Skill Learning

Reviewer signals → skill revision candidate

Graph: skill_learning_graph.py (skill-learning graph)

• Collects learning signals from reviewer edits and rejections across launches.
• Clusters the edit and rejection patterns and selects the impacted skills.
• Drafts a candidate revision to a repo SKILL.md and halts for human review.

Feeds Gate #3 — a human must approve before any repo SKILL.md is overwritten and its commit SHA recorded.

Review skill candidates