AgentOps AI 0.10.0

AgentOps AI 0.10.0

AgentOps AI is an evidence-grounded business intelligence and decision intelligence agent. It combines deterministic analytics, forecasting, anomaly detection, secured tool execution, MCP integration, and multi-step investigations to turn business questions into evidence-backed findings and recommendations.

Highlights

1. Evidence-grounded business intelligence

• KPI analysis across revenue, customers, sales, marketing, support, and product data.
• Every business claim is tied to structured evidence and calculation lineage.
• Clear distinction between observed, calculated, inferred, and recommended information.

2. Forecasting and anomaly detection

• Revenue, MRR, support volume, active customers, and product adoption forecasting.
• Rolling-origin backtesting with deterministic model selection.
• Multiple forecasting approaches including Naive, Seasonal Naive, Moving Average, Drift, and Damped Trend / ETS.
• Rolling z-score, IQR, and forecast-residual anomaly detection.

3. Decision intelligence investigations

• Multi-step investigations for business issues rather than only single-question analysis.
• Structured investigation plans, findings, drivers, evidence, and recommendations.
• Bounded investigation budgets to control tool calls, evidence, output, and execution time.
• Recommendations are generated only when supported by validated findings.

4. Agent reliability and evaluation

• Deterministic benchmark suite covering business questions, comparisons, rankings, refusals, numerical correctness, evidence grounding, and security.
• Final evaluation: 89/89 scenarios passed.
• Investigation evaluation: 77/77 scenarios passed.
• Critical scenarios and multi-seed evaluations passed.
• Claim/evidence identity validation prevents mismatched metrics, periods, dimensions, and comparisons.

5. Security and guardrails

• Bearer-token API authentication.
• Rate limiting and request lifecycle controls.
• Read-only, parameterized, allowlisted SQL execution.
• Shared secured execution path across the agent and MCP.
• Prompt-injection and unsafe-query defenses.
• Sensitive customer information is protected from inappropriate agent output.
• Structured JSON logging without questions, answers, tokens, secrets, or sensitive request data.

6. MCP integration

• Business analytics tools exposed through MCP.
• MCP calls use the same security and execution controls as the native agent.
• Direct-agent and MCP results were validated for parity.

7. API and UI

• FastAPI interface for questions and investigations.
• Streaming progress updates.
• Evidence, provenance, analysis trace, KPI cards, forecasts, anomalies, findings, and recommendations.
• Browser-session history without persistent user accounts.

8. Production hardening

• Dockerized API and UI.
• Non-root containers.
• Read-only filesystem configuration.
• Restricted container capabilities.
• Health/readiness endpoints.
• CI validation for tests, linting, formatting, type checking, evaluation, Docker builds, smoke tests, and graceful shutdown.

Evaluation

Final validation includes:

Area	Result
Standard evaluation benchmark	89/89
Investigation benchmark	77/77
Critical standard scenarios	13/13
Critical investigation scenarios	12/12
Security scenarios	32/32
Standard multi-seed evaluation	22/22
Investigation multi-seed evaluation	24/24
Final automated test suite	2,988 passed
Docker smoke tests	15/15

Scope and limitations

AgentOps AI currently operates on a synthetic Northwind Cloud SaaS dataset. Investigation drivers are based on accounting relationships, co-movement, anomalies, and associations rather than causal inference. Investigation workflows use a bounded set of deterministic templates, and the current deployment is designed as a single-process application with in-memory rate limits and session history.

This release focuses on a complete, evaluated, security-hardened demonstration of evidence-grounded business and decision intelligence.