Reproducible by design
Experiments are defined by versioned configuration, pinned data snapshots, and recorded model and prompt versions, so a result can be re-run and examined.
Research-stage platform
Market intelligence, engineered for rigorous research.
Market-Aware Agents (MAA) is an AI-native research platform from StreamPod. A Unified Broker Interface (UBI) over brokers’ MCP servers, a Python agent harness, and a deterministic policy engine work together so market research is reproducible, auditable, and clearly separated from any real-world action.
MAA is not an investment fund, advisor, or broker. It does not trade live, publishes no performance results, and does not accept investment capital.
Value proposition
Research you can inspect, repeat, and challenge.
Most AI-for-markets demos optimize for impressive output. MAA optimizes for something less glamorous and more useful: knowing exactly how a result was produced.
Agents propose, policy decides
Language-model agents produce structured research outputs. A deterministic policy engine — code, not a prompt — is the only component that can approve a downstream action.
Simulation is labeled as simulation
Historical simulations always carry their assumptions, cost models, and limitations. They are never presented as live results.
Contained and observable
Least-privilege access, immutable audit logs, and managed cloud services keep the research environment observable and contained.
Research capabilities
What the platform is being built to do.
Each capability is marked with its current stage. Nothing listed here is a live or production service.
Research
Market data ingestion
Normalize data from broker MCP servers (through UBI), data vendors, filings, and news into point-in-time datasets with provenance, quality checks, and as-of timestamps.
Research
Agent orchestration
Coordinate role-specialized agents — for example fundamentals, news, technical, and risk review — with structured, validated outputs.
Research
Research workflows
Move from hypothesis to data specification, experiment, and review with an experiment registry and written methodology.
Research
Policy and risk controls
Express limits and approval rules as deterministic code with human review, kill switches, and a complete decision log.
Research
Audit trail
Record prompts, tool calls, data versions, and decisions so any output can be traced back to its inputs.
Planned
Backtesting harness
Run historical simulations with explicit transaction-cost assumptions and out-of-sample validation.
Architecture overview
Models can suggest. Code decides.
Data flows left to right, and trust flows the opposite way. Model output is treated as untrusted input to a deterministic policy layer.
-
01
Market data sources
Brokers’ market-data services, data vendors, filings, and news — accessed under their terms.
External input
-
02
UBI and MCP servers
The Unified Broker Interface presents each broker’s MCP server, and other connectors, through one least-privilege interface.
Deterministic code
-
03
Python agent harness
Runs role-specialized agents with schema validation, run budgets, and full logging.
Model-driven
-
04
Policy engine
Rules-as-code. The only component allowed to approve any downstream action.
Deterministic code
-
05
Research outputs and audit log
Reports, experiments, and immutable records. In the research stage, outputs are documents — not orders.
Output
Component types
- External input
- Deterministic code
- Model-driven
- Output
Where we are
Research stage, stated plainly.
MAA is in the research stage. The work today is defining data schemas, prototyping the agent harness, and writing down methodology before running experiments at scale. Backtesting, paper trading, and any controlled deployment come later, and each stage is gated on the one before it. No dates are committed.
Research preview
Hear when research notes and previews are available.
Register your interest in preview materials. This is an information list only. It is not an investment product and nothing here is an offer or solicitation.