Build assistants that find evidence, use tools and produce work you can check. Learn through everyday applications, with trading as one specialist case study.
Your first result: A quotation with its source and an explicit review step.
Learn to build useful AI assistants through research, tool use, memory and document workflows. Complete examples and checkable results, with trading as one specialist case study.
Have a question?
Ask about anything The Agentic Trading Research Journal has covered. You get a short answer with links to the issues it came from.
Latest issue · November 9, 2026
LongMINT found that current memory systems struggled when facts were revised and questions required historical lookback or aggregation. A revision ledger keeps current and prior states answerable from the same dated evidence.
Read the latest issue37 published
Nov 8, 2026
A new 5,012-trajectory MCP benchmark found that unreachable goals often pushed agents toward false completion. A preflight feasibility card checks capabilities, tool errors and external state before an assistant says it is done.
Nov 7, 2026
OpenAI says Agent Builder will shut down on November 30, and its export does not guarantee that workflow behavior transfers unchanged. A two-column migration receipt checks both the requested result and every write.
Nov 6, 2026
Agent-commerce protocols can authorize and settle a purchase, but they do not decide who receives scarce inventory. A new market-microstructure survey shows why one small round has no allocation that helps everyone.
Nov 5, 2026
A new cross-framework study followed poisoned tool descriptions through real agent execution. Learn why model refusal is not a control and pin every approved tool definition before use.
Nov 4, 2026
A fresh benchmark exposes the gap between a polished agent-written paper and recorded work. Use integrity, completeness and a domain threshold before release.
Nov 3, 2026
A real PDF table exposes reversed headings and merged cells in MarkItDown 0.1.9. Inspect the source, see a limited repair and keep a checked extract before an assistant answers.
Nov 2, 2026
Quant-Arena's top score, highest return and smallest drawdown tell different stories. Read the real results, the overlooked setting distinction and what a sealed historical test still cannot prove.
Nov 1, 2026
Use MCP Inspector and the real filesystem server to discover tools, create a file, read it back and inspect an outside-folder denial. No model subscription is needed.
Oct 28, 2026
Separate standing AGENTS.md instructions from today's coding task. Finish two editable notes, inspect stale-file and ordering mistakes, and run a complete optional local example.
Oct 27, 2026
A working DOI can still point to the wrong reference. Use a real lookup, a corrected citation and an editable review note to separate publication identity from claim support.
Oct 21, 2026
Use five fictional records to produce a readable decision brief, an eight-row claim table and two visible unknowns, then reject stale, altered and invented claims.
Oct 20, 2026
Build a local document worker with saved progress, a pause control and a lifetime work allowance. Test restarts and incomplete jobs without calling a model or live service.
Oct 19, 2026
Build a complete offline scorecard for workshop request routing, compare mistakes and human effort, and prevent missing results or unknown costs from becoming a success claim.
Oct 18, 2026
Preview a complete notice, bind a local decision to the reviewed values, and test rejection, changed input and repeat saves with a complete Python exercise.
Oct 17, 2026
Create a persistent project notebook in Python and SQLite, reject stale updates, and make expired facts trigger a recheck instead of reviving an old answer.
Oct 16, 2026
Build a tested CSV stock lookup, reject malformed tool requests, and distinguish an unknown product from zero stock without reserving or changing anything.
Oct 15, 2026
Create a cited planning brief from approved notes, reject invented references and altered quotations, and make missing evidence and human review visible.
Oct 14, 2026
Build an offline partial-fill reconciliation report that waits for complete cancellation evidence, counts each fill once and preserves the remaining quantity before planning a replacement.
Oct 13, 2026
Build a fail-closed explorer fallback that recovers from technical failures without overriding an unverified response or weakening token safety policy.
Oct 12, 2026
Build an offline candidate-trace report that separates explicit blocks from missing evidence, handles duplicate events and delayed receipts, and rejects contradictory histories.
Oct 11, 2026
Build a synthetic size-aware exit calculator that separates chart gains from net proceeds, preserves unsold inventory and rejects stale or incomplete evidence.
Oct 10, 2026
Predeclare a minimum-evidence gate, preserve unresolved folds and refuse promotion when a favorable aggregate is too small, incomplete or concentrated.
Oct 9, 2026
Repeat a frozen ranking comparison across equal chronological folds, preserving unavailable labels, costs and coverage before judging consistency.
Oct 8, 2026
Compare two frozen rankers without letting later labels, missing outcomes or inconsistent costs leak into the result.
Oct 7, 2026
Keep open, receipt-pending and incomplete-cost trades out of win/loss labels, while reporting exactly how much of the cohort has verified outcomes.
Oct 6, 2026
A companion guide for freezing inputs, checking decision-time availability and replaying a comparison without inventing trading results.
Oct 6, 2026
Hold evidence fixed, separate admission changes from scoring changes, and reproduce a controlled comparison without inventing trading outcomes.
Oct 5, 2026
Reproduce a market-data overwrite bug from our research changes, preserve exact identity, and test freshness before a model sees the observations.
Oct 4, 2026
Separate quote age, response age and decision-time availability so an agent and its comparison rules see the same admissible evidence.
Oct 3, 2026
Continue our automation experiment at the trading boundary: distinguish duplicate, missing, conflicting and partial execution evidence before any retry.
Sep 26, 2026
Record job states, distinguish retryable failure from an uncertain outcome, and prove that a repeated run does not duplicate a completed local result.
Sep 26, 2026
Separate generated suggestions from validated decisions, reject unsupported evidence, and make an agent workflow inspectable.
Sep 26, 2026
Separate rule selection from later evaluation, keep a baseline, and publish an honest result when a challenger fails.
Sep 26, 2026
Build a transparent paper ledger, distinguish gross from net outcomes, and see why costs can overturn a promising-looking result.
Sep 26, 2026
Create an isolated workspace, generate a small synthetic dataset, and leave a record another reader can reproduce.
Sep 25, 2026
An introduction to our research series and the evidence needed to tell learning from a lucky trading result.
Follow The Agentic Trading Research Journal by email or keep up in your favorite feed reader.
Bring readers who share your interests and track your publication's referral milestones.
Your referral rewards ↗If this was useful, you can leave a tip. It goes straight to the writer.
Published on Publish Haven. Start your own newsletter free →
publish