Imported from AmaanBarmare/resolvr (
AGENTS.md). Install upstream withnpx skills add AmaanBarmare/resolvr. Copyright stays with the author.
AGENTS.md — Building Agent A and Agent B
The two agents are Python programs that run inside Veris. Each agent takes a scenario, calls mock tools, and produces a structured recommendation with named assumptions.
What Agents Need to Do
Each agent must:
- Receive the TechFlow scenario as a prompt
- Call its mock tools (Salesforce for A, finance model for B)
- Reason about the tool outputs
- Return a structured recommendation with every number explicitly named and sourced
The named numbers are critical — the Forensics Agent reads the Veris transcript and extracts them. If the agent just says "growth looks good" without citing specific numbers, forensics produces nothing useful.
Agent A — Revenue Agent
Role: Optimize for pipeline velocity and growth. Bullish on APAC opportunity.
Tools it calls:
salesforce.getOpportunities→ pipeline value, deal count, stage breakdownforecast.getCloseRate→ historical close rate from internal CRMforecast.getARRProjection→ projected ARR if deals close
System prompt:
You are a Revenue Strategy Agent for a Series A B2B SaaS company.
Your job is to analyze pipeline data and recommend hiring decisions
based on revenue opportunity.
You have access to:
- salesforce: to query the current deal pipeline
- forecast: to get close rates and ARR projections
When making recommendations, always cite:
- The exact pipeline value you found
- The exact close rate you used and its source
- The projected ARR impact
- The specific number of engineers you recommend hiring
Output your final recommendation as JSON with this structure:
{
"recommendation": "string — your hiring recommendation",
"assumptions": [
{
"variable": "machine_readable_key",
"value": "value with units",
"source": "where you got this"
}
],
"reasoning": "your reasoning chain"
}
File: veris/agent_a/revenue_agent/main.py
import os
import json
import httpx
from fastapi import FastAPI, Request
app = FastAPI()
BASETEN_API_KEY = os.getenv("BASETEN_API_KEY")
BASETEN_MODEL_ID = os.getenv("BASETEN_MODEL_ID")
SYSTEM_PROMPT = """You are a Revenue Strategy Agent....""" # full prompt above
TOOLS = [
{
"name": "salesforce_get_opportunities",
"description": "Get current deal pipeline from Salesforce",
"input_schema": {
"type": "object",
"properties": {
"stage": {"type": "string", "description": "Deal stage filter"}
}
}
},
{
"name": "forecast_get_close_rate",
"description": "Get historical close rate from internal CRM",
"input_schema": {
"type": "object",
"properties": {
"source": {"type": "string"},
"period": {"type": "string"}
}
}
},
{
"name": "forecast_get_arr_projection",
"description": "Project ARR based on pipeline and close rate",
"input_schema": {
"type": "object",
"properties": {
"pipeline_value": {"type": "number"},
"close_rate": {"type": "number"}
}
}
}
]
@app.post("/")
async def handle_message(request: Request):
body = await request.json()
user_message = body.get("message", "")
# Call Claude via Baseten with tool use
response = await call_claude_with_tools(user_message)
return {"response": response}
async def call_claude_with_tools(scenario: str) -> str:
# Baseten inference call
async with httpx.AsyncClient() as client:
resp = await client.post(
f"https://model-{BASETEN_MODEL_ID}.api.baseten.co/production/predict",
headers={"Authorization": f"Api-Key {BASETEN_API_KEY}"},
json={
"messages": [{"role": "user", "content": scenario}],
"system": SYSTEM_PROMPT,
"tools": TOOLS,
"max_tokens": 2000
}
)
# Handle tool use loop
# Return final JSON recommendation
...Agent B — Risk Agent
Role: Optimize for survival. Protect runway. Conservative on macro.
Tools it calls:
finance.getBurnRate→ current monthly burnfinance.getRunwayMonths→ months of runway at current burnfinance.getHireCostImpact→ how much each hire adds to monthly burnmacro.getMarketOutlook→ macro enterprise SaaS conditions
System prompt:
You are a Risk Management Agent for a Series A B2B SaaS company.
Your job is to analyze financial risk and make conservative
headcount recommendations based on runway preservation.
You have access to:
- finance: to query burn rate, runway, and cost projections
- macro: to get market condition indicators
When making recommendations, always cite:
- The exact burn rate you found
- The exact runway in months
- The cost impact of each hire on monthly burn
- The macro market conditions you factored in
Output your final recommendation as JSON with this structure:
{
"recommendation": "string — your hiring recommendation",
"assumptions": [
{
"variable": "machine_readable_key",
"value": "value with units",
"source": "where you got this"
}
],
"reasoning": "your reasoning chain"
}
Veris Configuration
Agent A — veris/agent_a/veris.yaml
services:
- name: salesforce
dns_aliases:
- api.salesforce.com
- api.salesforce.mock
- name: forecast-api
dns_aliases:
- forecast.internal
- forecast.techflow.mock
actor:
channels:
- type: http
url: http://localhost:8001
method: POST
headers:
Content-Type: application/json
request:
message_field: message
session_field: session_id
response:
type: json
message_field: response
agent:
code_path: /agent
entry_point: python -m revenue_agent.main
port: 8001
environment:
BASETEN_API_KEY: "" # set via: veris env vars set BASETEN_API_KEY=xxx --secret
BASETEN_MODEL_ID: ""Agent B — veris/agent_b/veris.yaml
services:
- name: finance-model
dns_aliases:
- finance.internal
- finance.techflow.mock
- name: macro-api
dns_aliases:
- macro.internal
- macro.research.mock
actor:
channels:
- type: http
url: http://localhost:8002
method: POST
headers:
Content-Type: application/json
request:
message_field: message
session_field: session_id
response:
type: json
message_field: response
agent:
code_path: /agent
entry_point: python -m risk_agent.main
port: 8002
environment:
BASETEN_API_KEY: ""
BASETEN_MODEL_ID: ""Agent A — veris/agent_a/Dockerfile
FROM python:3.11-slim
WORKDIR /agent
COPY requirements.txt .
RUN pip install -r requirements.txt
COPY revenue_agent/ revenue_agent/
CMD ["python", "-m", "revenue_agent.main"]Agent B — veris/agent_b/Dockerfile
FROM python:3.11-slim
WORKDIR /agent
COPY requirements.txt .
RUN pip install -r requirements.txt
COPY risk_agent/ risk_agent/
CMD ["python", "-m", "risk_agent.main"]What a Good Transcript Looks Like
After running Veris, the transcript JSON for Agent A should look like this. If it doesn't contain named numbers with sources, the forensics agent will fail.
{
"agent": "revenue_agent",
"scenario": "techflow_hire_decision",
"tool_calls": [
{
"tool": "salesforce.getOpportunities",
"input": {"stage": "qualified"},
"output": {
"pipeline_value": 4200000,
"deal_count": 23,
"avg_deal_size": 182608,
"top_region": "APAC"
}
},
{
"tool": "forecast.getCloseRate",
"input": {"source": "internal_crm", "period": "Q2_2026"},
"output": {
"close_rate": 0.34,
"period": "Q2_2026",
"source": "internal_salesforce_crm",
"confidence": "high"
}
},
{
"tool": "forecast.getARRProjection",
"input": {"pipeline_value": 4200000, "close_rate": 0.34},
"output": {
"projected_arr": 1428000,
"timeline_months": 6
}
}
],
"reasoning_steps": [
"Current APAC pipeline: $4.2M across 23 deals.",
"Internal CRM shows 34% close rate for Q2 2026.",
"At 34% close rate: $4.2M × 0.34 = $1.43M new ARR.",
"To deliver APAC roadmap in 6 months, need 12 engineers.",
"Revenue justifies headcount investment."
],
"final_recommendation": "Hire 12 engineers immediately. Pipeline velocity demands it.",
"key_assumptions": {
"close_rate": {"value": "34%", "source": "internal_crm_q2"},
"pipeline_value": {"value": "$4.2M", "source": "salesforce"},
"engineers_needed": {"value": "12", "source": "engineering_estimate"}
}
}Mock Transcripts for Development
If Veris isn't ready, generate mock transcripts for local development:
cd backend
python scripts/generate_mock_transcripts.py
# Creates backend/data/transcript_a.json and transcript_b.jsonThe mock transcripts are realistic enough for the forensics agent to work on. Use them during development, then replace with real Veris transcripts before the demo.