Build an investment research Agent
An investment research Agent can stay available in AstraBox to collect permitted company data, run repeatable analysis, and save reports in its cloud workspace. The work continues in the cloud after the developer's laptop disconnects, and the same Agent can be started from the web, a schedule, a signed webhook, or a messaging platform.
| Capability | What it enables |
|---|---|
| Research method | A prompt, Skill, or Plugin can define sourcing, calculations, review, and output requirements. |
| Current data | A remote MCP server or an Agent program's web capability can reach approved data sources. |
| Analysis | The Agent program can run scripts and use files in the Session workspace. |
| Automation | A Deployment can start the Agent on a schedule or when an external event arrives. |
Start with the research task
A simple research Agent may need only a model and a clear prompt. Add MCP servers, Skills, Plugins, credentials, and wider network access only when the task needs them.
Open Management console → Agents and create or edit an Agent. A useful research prompt can require the Agent to:
- cite the source and reporting date for every material fact;
- separate reported facts, calculations, and interpretation;
- show formulas, units, currencies, and period definitions;
- explain missing or conflicting data; and
- save useful tables, scripts, and reports in the workspace.
Choose an Environment that runs the required Agent program and permits the connections the research sources need. An Environment determines the sandbox, network access, and model connection used for the work.
Connect research sources
Add a remote MCP server when a licensed filing, market-data, document, or internal service provides one. The server remains remote whether its definition is entered directly on the Agent or supplied by a Plugin.
When the source requires a credential, store it in a Credential Vault and use protected delivery for the source's exact destination. If the Environment uses limited networking, allow the remote host. The Agent should only access data that the operator and user are authorized to use.
See Configure MCP servers, Plugins, and Skills, Credential Vaults, and Protect credentials used by Agents.
Add a repeatable research method
A Skill can provide a focused procedure and supporting files. A Plugin can package Skills, commands, and MCP definitions. Both are optional: the selected Agent program keeps its native capabilities when neither is configured.
Use a reviewed source and pin Git-based extensions to a reviewed revision. The research method can define how to choose comparable periods, handle restatements, calculate metrics, cite evidence, and format the final report.
Run and inspect the work
Start the Agent and give it a concrete task, for example:
Compare reported segment revenue for the last three fiscal years. Cite the filing and reporting date behind every value, explain classification changes, save the extraction script, and write the result to
segments.csv.
The conversation and its work belong to an AstraBox Session. The Agent can read
and write files under /workspace; the web interface can preview and download
those files. Keeping the script, intermediate data, and report together makes
the result easier to reproduce and review.
Run the research automatically
Use a scheduled Deployment for recurring monitoring. Use a signed webhook when another system should start the Agent after an event, or connect a messaging platform when people should request and receive research from chat.
Each scheduled or webhook invocation creates a new Session. A messaging platform starts or continues the Session mapped to its external conversation.
See Automate Agent runs for these trigger options.
Review the result
Check every cited source, date, calculation, unit, and saved file before using the output. Confirm that comparable companies and periods use the same basis, and treat missing data explicitly. Generated research is an aid to analysis, not personalized investment advice.