Dynamic Enterprise Agent

Dynamic Enterprise AI Agent Builder — Strategic Orchestrator

Build a fully functional, deployable AI agent in real time — while the conversation is still happening. No post-meeting spec docs. No multi-week discovery. The agent exists before you hang up.

Core Principle

Listen → Decode → Build → Ship.

The moment someone describes a problem, this skill orchestrates four specialist skills in sequence: extract what is needed, map it to the right platform and industry context, generate working code and configuration, and produce a deployment-ready agent. Every output is actionable, not a proposal.

Pipeline Overview

Orchestration Instructions

Entry Points

Based on what the user provides, enter the pipeline at the right stage:

Stage 1 — Listen

Invoke: enterprise-signal-listener

Purpose: Capture signals from the live conversation. Do not interrupt. Do not ask for structured input. Extract platform, industry, pain trigger, actor, frequency, and desired outcome from natural speech.

Ask at most 2 clarifying questions. Stop as soon as enough signal exists to decode.

Gate: Signal capture is complete when spoken_problem, platform_hint, and desired_outcome are all populated. Proceed immediately — do not wait for a perfect intake.

Output → Decode: Pass the completed intake YAML to enterprise-problem-decoder.

Stage 2 — Decode

Invoke: enterprise-problem-decoder

Purpose: Map intake signals to a precise agent specification. Resolve the exact platform entry point, apply industry compliance constraints automatically, and classify the agent type (Monitoring, Triage, Automation, Decision-Support, or Orchestrator).

Gate: Agent spec is complete when platform, agent type, trigger, and desired outcome are all resolved. Flag open questions but do not block — builders resolve ambiguity.

Output → Build: Pass the agent_spec YAML to enterprise-agent-builder.

Stage 3 — Build

Invoke: enterprise-agent-builder

Purpose: Design the complete agent architecture (blueprint + data flow diagram) and generate production-ready Python code using the Anthropic Claude API, with platform-specific adaptations for ServiceNow (JavaScript), Salesforce (Apex), and Snowflake (Snowpark) where required.

All guardrails are embedded in the generated code. No irreversible actions without human approval. Audit trail on every tool call. Credentials via secrets manager only.

Gate: Build is complete when the blueprint, data flow diagram, and runnable code are produced. The code must be functional as written — not a skeleton requiring rewrite.

Output → Ship: Pass the blueprint, code, and agent spec to enterprise-agent-shipper.

Stage 4 — Ship

Invoke: enterprise-agent-shipper

Purpose: Produce the full deployment package: environment variable configuration, secrets manager integration, deployment target selection, least-privilege IAM policy, a live demo script, and a ship checklist. The agent is deployment-ready when this stage completes.

Gate: Ship is complete when the demo script and ship checklist are produced and all open items on the checklist are documented (not necessarily resolved — but known).

Output: Deployable agent package. Hand to the client or engineering team with the checklist.

Quick-Build Matrix

Match any problem heard in conversation to a pre-configured starting point. Feed the row into enterprise-signal-listener or skip directly to enterprise-problem-decoder:

See references/agent-templates.md for full template implementations.

Non-Negotiable Guardrails

These apply to every agent produced by this pipeline. They are embedded by enterprise-agent-builder and verified by enterprise-agent-shipper. Never remove them.

• No irreversible action without human gate — delete, bulk-update, financial transaction, patient record change, or security policy change always requires request_human_approval first.
• Audit trail on every action — every tool call logged with timestamp, actor identity, input, and output.
• Credential hygiene — all secrets via environment variables or secrets manager. Zero hardcoded credentials.
• Scope-limited API credentials — read-heavy with surgical write scope. No admin credentials.
• PII/PHI handling — if healthcare or life sciences, de-identify before any LLM call not within a compliant boundary.
• Rate limiting — exponential backoff and per-minute rate limits on all platform API calls.
• Graceful degradation — if platform API is unreachable, fail safe: log and alert a human. Never guess or hallucinate data.

Regulated-Industry Overrides

Pipeline Output Summary