Deloitte sponsored practicum · multi-agent
Consulting Workplan Agent
129
automated tests across agents, API and exports
Consulting deliverables are heavily cross-referenced: every risk points at work packages, every work package has an owner, every owner appears in the RACI. The design question was how to let LLM agents generate these artifacts while keeping the references consistent enough for a human reviewer to trust.
Client brief
Brief parser
confirm open questions
WBS agent
RACI agent
RAID agent
Validators
cross-artifact checks
Human review
revise + version
DOCX / XLSX export
Design
- An orchestrator owns session state, artifact versions and persistence. Six stateless specialist agents (brief parser, WBS, RACI, RAID, status report, revision) each return a validated Pydantic model, not free text.
- Open questions from the brief are surfaced to the human before generation; feedback during review goes to a revision agent that produces a new artifact version with a diff.
- Everything runs offline with a mock LLM provider, which is what the demo and most of the tests use.
Keeping artifacts consistent
- Reconciliation and validators check that every leaf WBS package appears in the RACI, that WBS owners hold a Responsible or Accountable role, and that RAID owners and WBS links resolve to real entries.
- A generated quality-gates checklist tells the reviewer what a human still has to verify, and a failure-mode table maps each known failure (orphaned risk links, mis-aligned RACI, wrong RAG status) to its mitigation.
LLM plumbing
- A routing provider dispatches each call to Anthropic, OpenAI, Gemini or DeepSeek by the configured model, and falls back to another available provider when a key is missing.
- For Claude, prompt-cache warm-up primes the cached prefix (parsed brief, WBS, RACI) before the expensive WBS and RAID calls so repeated tokens are not billed at full price.
- Exports cover JSON, Markdown, DOCX and XLSX; a template-sync script pulls the firm's layout so deliverables match house style.
Stack
PythonPydanticFastAPIClaudeGPTGeminiDeepSeekpython-docxpytest