Claude Managed Agents

Claude Managed Agents remain SABLE's reasoning and writing layer. The reliability architecture does not remove or replace the SABLE managed agent: Trigger.dev owns the clock, connector refresh, readiness gate, validation, and one-time delivery; a temporary managed-agent session produces the structured draft from scoped Supabase evidence.

Trigger schedule -> connector refresh -> readiness gate -> SABLE CMA draft
                 -> evidence validation -> deterministic render -> persist -> deliver once

The managed agent cannot deliver a scheduled board directly. Report sessions use a read-only spine credential and a fixed session override. The response must be JSON, bind to the Trigger run/date/kind, and attach evidence to every item. Trigger rejects stale, malformed, or ungrounded drafts.

The interactive agent, report staging, learning loop, and Trigger orchestrator each use independently generated credentials. No restricted credential is formed by appending a suffix to a more powerful secret. Only Trigger receives the credential that can persist or deliver a board.

Agent configs live in:

sable-agents-demo/agents

Agents

Agent	File	Role
sable-core	agents/sable-core.yaml	Morning board narrative, synthesis, action-code drafts
fathom-core	agents/fathom-core.yaml	Meeting transcript extraction in the original demo
learning-loop	agents/learning-loop.yaml	Converts corrections into proposed prompt/skill candidates
asana-agent	agents/asana-agent.yaml	Asana routing/housekeeping worker config using hosted Asana MCP

Memory Store Demo Path

The original demo stores a fictional spine in a CMA memory store:

• /spine/*.json
• /skills/*.md
• /prompts/*.md
• /board/template.html
• /out/board.html
• /inbox/*.md

This path still supports the older "drop transcript, render board, draft follow-up" demo.

Supabase Production Direction

The current production direction is:

• Supabase stores canonical connector truth.
• Edge Functions ingest connector data.
• Vercel API reads Supabase.
• Managed Agents operate on scoped data and write structured, evidence-bound drafts.
• Trigger.dev is the single production scheduler for morning and EOD boards.
• Supabase persists connector checkpoints, activity evidence, board drafts, and delivery state.
• The learning loop runs separately after EOD and can only propose human-reviewed prompt rules.

The live base prompt and human-approved learned rules remain unchanged by report orchestration. The report and learning contracts are temporary session overrides. Lindsey remains the prompt design authority for promotion or rollback.

The older Vercel /api/fathom-webhook route remains demo-only. Canonical Fathom ingestion is the Supabase fathom-webhook Edge Function.