Skip to main content
The harness is everything around the model: what it may do, what it must prove, who checks it, and when it runs again. It exists so that you can trust the work without redoing it.

The research agent

The research agent plans the work, writes Python against Itoflow’s quant library, runs it in an isolated sandbox, and answers with the code behind each result.
  • Code, not a menu of tools. The agent writes one script that can fetch data, run a test and measure risk together, rather than pick from a long list of separate tools. You can read and rerun the code afterward.
  • An isolated sandbox. Each chat has its own private workspace. The sandbox holds no database credentials and no tokens; it reaches data only through fixed operations, scoped to the owner, that Itoflow runs outside the sandbox.
  • A full record. A finished turn holds a dashboard (text, tables, charts, formulas and method diagrams), progress updates, the agent’s working notes, the review, and every file it wrote. What a finished turn contains.
  • It asks before it assumes. Choices about your objective, risk, capital or permissions need your answer before the research that depends on them starts.
  • It says what it couldn’t do. If a core data source fails, the agent stops that analysis and explains why. If a minor input is missing, it leaves it out and states the gap in the answer.
  • You stay in control of the turn. From your assistant you can start a turn, follow it, cancel it, or send it again safely: the same message ID is the same turn.
What our quant-research agent actually does — and where it still needs a human shows two real runs, including one where the agent tested two signals out of sample, found no edge, and said so rather than invert the signals or switch universes.

The research contract

Before any research runs, the agent turns your request into a research contract: the question, the universe, the data and dates, the methods it will try, how it will choose between them, and the evidence the answer must carry. The system checks the contract and sends back one that is incomplete or badly formed, before any work starts. While it works, each number the agent reports needs a proof that the system checks: Proofs cover numbers and files. Judgement, such as whether a method suits the question or a result is plausible, is the reviewer’s job. Overfitting checks are part of the engine: the probability of backtest overfitting, the deflated Sharpe ratio and false-discovery-rate control. The contract can require them as evidence. If the agent needs a method the library doesn’t offer, it may declare a custom method with a short description, and the reviewer checks that no library method fits. The contract has limits, and we say so: it moves ambiguity to the start of the work rather than remove it, and a portfolio that passes every check is not therefore a profitable one. How Itoflow turns investment intent into a machine-checked contract explains the design.

Independent review

When the agent builds a research contract, Itoflow starts a second agent to review the work at two points: the plan, before research begins, and the answer, before it reaches you.
  • Itoflow starts it, not the research agent. The research agent can’t skip the review or narrow what it covers.
  • It reads the work itself. The reviewer reads the actual files and results, not the research agent’s summary of them.
  • It doesn’t change the work. In chat research it has read-only access to the workspace and approved research tools; a strategy’s reviewer may also keep its own notes. It can’t edit the research agent’s files, trade, ask you questions or start other agents.
  • It gives a verdict. It accepts the work or sends it back with what to fix. The reviewer normally follows repairs in the same conversation. If a strategy review runs out of context, Itoflow retries with the current request and work; the strategy’s saved review notes remain.
  • It runs on your model. The reviewer uses the model you chose for the run, so it can share that model’s blind spots. That is why exact rules, such as the contract checks, position limits and trade approval, sit in code rather than with either agent.
  • When work goes on without acceptance. If the review itself keeps failing, or concerns are still open after 10 rounds, Itoflow records which and the work continues, and the system’s own checks and portfolio limits still apply.
In published examples, review caught a position cap the agent had made up, survivorship and look-ahead bias in a historical test, and a call to a function that doesn’t exist. On 7 matched synthetic research tasks, with one run each, review raised the mean score from 0.568 to 0.693 on a 0 to 1 scale; six tasks improved and one got worse. How we use a second AI agent to review investment research, July 30, 2026, gives the method, the failure and an earlier design that didn’t work.

Agents that keep working

Monitors and strategies are agents that run for weeks or months. Each wakes, does its work, records what it found or decided, and sleeps until the next run. Monitors watch a list of holdings or a whole fund, such as IVV, and report what changed on a schedule you set.
  • The engine computes the figures before the agent starts. The agent picks what matters, writes the findings and explains them.
  • Before a report goes out, its figures are checked against the engine’s output. A figure the engine can’t produce shows as “not known”.
  • Asking for a report while one is queued joins that run rather than start a second.
Strategies run a paper or live portfolio under a mandate you write.
  • At each run the agent works as of that date and reads only data that existed then.
  • It decides when to wake next: on a date, on a move in a held name, or when a drawdown limit is hit.
  • It records each decision with its weights, its reasoning and its own stated confidence.
  • A strategy the agent drafts starts paused, and you turn it on. Paper portfolios can then trade automatically. Live portfolios wait until a person approves the exact list of trades, named by its ID and a hash of its contents.
  • Strategy agents meet the same evidence standard as chat. They decide within a standing mandate rather than ask you each time.
Monitors and Strategies cover setup and the API.

Skills, instructions and memory

Agents don’t start from nothing.
  • Skills. Agents read written procedures for the job, such as how to run a strategy, design wake conditions, research an allocation or read historical factor risk. Skills are plain files, reviewed and changed like code, so a change in behavior is a change you can read.
  • Guidance when it applies. Help arrives in tool results at the moment it is needed, rather than in one long prompt at the start.
  • Memory. Strategy agents keep notes between runs, such as lessons, research and hand-off notes, and read them each time they wake.

Models

You choose the model for each chat, monitor and strategy. A run uses the model saved when it starts, for the agent and its review, through every retry. If your plan stops covering that model, runs stop and ask you to choose; Itoflow never switches model without telling you. Chat and monitors can also run on your own ChatGPT subscription. See Models and plans.

Further reading