agent-specs

Agent specs — source of truth for initial Foundry agent creation

Flagship agents at a glance

The flagship scenario (Sales Research & Personalised Outreach) ships with five Foundry agents. Detailed system instructions live in the per-agent files in this directory; a partner-facing summary first:

Agent	Role	Output	HITL / tool posture
accel-sales-research-supervisor	Orchestrator — routes the scenario across the four workers and composes their outputs	Composed account research + outreach payload	No tools of its own; downstream tool calls inherit each worker's HITL policy
accel-account-planner	Builds the account brief: company, segment, signals, decision-makers	Structured account profile (citations required)	Read-only retrieval; no HITL
accel-icp-fit-analyst	Scores ICP fit + maps to a tier recommendation	Fit score + recommended play	Read-only retrieval; no HITL
accel-competitive-context	Surfaces competitive context + cloud footprint signals	Competitive notes + footprint signals	Read-only retrieval; no HITL
accel-outreach-personalizer	Drafts the personalised outreach + invokes side-effect tools (CRM write, send email)	Outreach copy + tool-call results	HITL required on every side-effect tool (crm_write_contact, send_email use HITL_POLICY = "always")

Use the per-agent files below for the actual system instructions; everything else on this page is bootstrap mechanics.

---

Bootstrap mechanics

src/bootstrap.py reads one Markdown file per agent from this directory and creates or updates the corresponding agent in the Foundry project at azd up / azd postprovision time.

File format

# Agent: <agent_name>

**Pattern:** <one-line description of what shape this agent fills>

## Instructions
<system instructions the agent runs with>

The model deployment is NOT declared here — every agent runs against the single model deployed by infra/modules/foundry.bicep (captured in the AZURE_AI_FOUNDRY_MODEL output and read by src/bootstrap.py at FastAPI startup). This keeps infra as the source of truth and prevents specs drifting away from what azd up actually provisions. The accelerator lint fails if any spec file contains a **Model:** field.

Important

Authoring source of truth: the .md files in this directory. Runtime location: Foundry portal — src/bootstrap.py syncs each spec verbatim to the matching Foundry agent at FastAPI startup on every azd up / azd deploy / revision restart. The portal holds the runtime copy between deploys but is not a durable authoring surface — manual portal edits are transient and will be overwritten on the next sync.

The supported authoring loop is: edit the .md → azd deploy → re-run evals. The supported rollback path is: git revert the spec file → azd deploy.

If an engagement intentionally opts out of bootstrap sync (sets BOOTSTRAP_SKIP=1 so the portal becomes the authoring surface for that deployment), the deviation must be recorded in the handover packet's "Customer-specific deviations" section.

Do NOT reference these .md files at runtime. Do NOT import from them.

Flagship agents (5)

The flagship ships one supervisor plus four workers. Each agent's .md file in this directory is the source of truth for its instructions; the table at the top of this page is the partner-facing summary. Add new agents to a scaffolded scenario with python scripts/scaffold-agent.py (writes a matching spec stub).