ETF Agent · NotesBuild financial research agents from scratch. No Agent framework required.
Learn long-running tasks, multi-agent collaboration and verifiable research delivery through fund research. Clear principles. Runnable source code.
Learn long-running tasks, multi-agent collaboration and verifiable research delivery through fund research. Clear principles. Runnable source code.
READ WITH A PURPOSE
Browse by category, or choose a topic above.
From the first example to the six-layer panorama.
Architecture, design decisions and system thinking.
Research and practice from the finance tag.
From requirements and implementation to testing and delivery.
New articles and content changes, based on source history. Prose whitespace-only edits are skipped. Dates: UTC+8.
When does another expert earn its cost? Follow one fund-research question through general Children, specialist methods, Reviewer and Bull/Bear exchanges, tracing evidence into independent contexts and results back to Main while weighing repeated preparation against verifiable contributions.
Run the fixed offline example to inspect source identity, tool execution, partial writes, and independent file readback, then separate known facts from gaps in a real-model rejection.
Start with the background and purpose of specifications, then follow file ingestion through principles, design, templates, behavior, verification, archiving and template maintenance—without copying project-specific ceremony.
How does a follow-up become a legitimate new task? A practical guide to message idempotency, conversation history, clarification binding, and explicit Agent Run continuation.
What changes when ordinary research becomes a long-running task? Follow a three-fund comparison through concrete integration steps for request assembly, budgets and finalization, material rereading, checkpoints, and report revisions, with links to the existing implementation guides.
A field-tested architecture case study of Core API, Product API, and Background Worker: measured lifecycle isolation, foreground capacity handoff, crash recovery, and local transport identity, without claiming production SLA gains.
Follow a fund-comparison conversation down to tasks, runs, model turns, request attempts, and tool actions. See which identities continue through replies, process recovery, and scheduled execution—and use those distinctions to design and debug an agent.