Algo Agent

Professional

The Algo Agent is an AI assistant embedded in the Algo Lab. In Agent mode it does not run a single turn and stop — it runs an autonomous multi-step research loop: planning, running backtests, proposing strategy variants, evaluating them on holdout data, and committing verified candidates back into your Case Study.

This page covers what that means in practice and where you stay in control.

The drawer stays beside the active Lab workspace, so the conversation and its evidence remain visible while you inspect the chart or report.

How to start an Algo Agent run

There is no separate “Run loop” button. The autonomy is the mode.

1. Open the Algo Agent rail on the right side of Algo Lab.

2. Make sure the mode selector shows “Agent” (below the chat input). Other modes — Ask (conversational) and Research (focused one-shot) — do not start the autonomous loop.

3. Type your research goal and press Send. Examples:

   • “Take this AAPL strategy and test one bounded RSI-filter hypothesis against the current baseline.”
   • “Evaluate whether a stricter RSI threshold improves the locked objective without violating the drawdown and trade-count gates.”
   • “Test three different stop-loss approaches on the active strategy and pick the best one.”

4. The agent runs autonomously. You’ll see “Thinking” / “Working” indicators, tool calls (analyse losing trades, propose candidate, evaluate, verify, commit), and a running step counter.

5. The loop stops on its own when any of these is true:

   • The agent commits a verified candidate
   • A turn produces no new research evidence
   • The Case Study ledger stops advancing (no new evidence is being recorded)
   • The configured budget (steps / tokens / runtime) is exhausted
   • The mission reaches a terminal state (e.g. ask-user required)
   • You hit Stop

Continuing a run

When the loop stops on a non-terminal condition (most commonly “no new evidence this turn”), the composer hints: “(or type ‘continue’ to keep researching)”.

Recognised continuation verbs:

• continue
• keep going
• next
• next step
• next iteration
• proceed
• go on
• more
• keep researching
• continue the loop
• continue research

Type any of these as a chat message in Agent mode and the agent picks up where it left off, using the existing Case Study state.

Chat verbs for staged candidates

When the agent commits a verified candidate, it stages the candidate’s source into your editor as a diff view. You’ll see an alert banner at the top of the page and a side-by-side diff in the editor tab.

You can accept or reject the staged candidate three ways:

1. Click Accept or Reject in the banner
2. Click ✓ Accept All or ✗ Reject All in the diff editor’s toolbar
3. Type a verb in the chat:

Accept verbs	Reject verbs
accept	reject
apply	discard
ship it	cancel
land it	drop it
commit it	undo
use it	never mind
yes	no

All three paths run the same five-step accept protocol (claim → re-check → capture → write → finalize). Read more in Engine vs TradingView about why you should still review the candidate’s source before saving it.

What “verifier-pass” means (and doesn’t)

The agent labels a candidate verifier-pass when it satisfies our statistical gates:

• Beat the baseline on the locked objective metric (e.g. netProfit)
• Beat the baseline on both the training slice and the held-out slice of historical bars
• Pass minimum trade-count, stability, and edge-margin checks
• Holdout block present with holdoutPassed: true

This means the candidate met the configured historical evidence gates for that run. It does not mean:

• The strategy will be profitable on live capital
• The strategy will reproduce on TradingView’s engine (more here)
• The strategy is robust to regime change after the historical sample
• You can skip independent forward validation

Treat verifier-pass as “worth your serious review”, not as a buy signal.

What it costs

The Algo Agent consumes hosted AI credits while it runs. Usage depends on:

• Model selected for the session
• Prompt size — your strategy’s source code, the Case Study state, prior turn history, all flow into the prompt each turn
• Tool output size — evaluate_strategy_candidate and verify_strategy_candidate return substantial JSON; those bytes count
• How long the loop runs before stopping

The app shows the available credit balance and stops the run when the configured budget or available credits are exhausted. Because source length, evidence context, tool output, and model rates vary, the documentation does not promise a fixed cost per research loop.

Watch the live balance

Use Account → Billing for the balance, usage history, reset date, and any purchase or request action available to your account. Model rates and research context can change, so the current product state is more reliable than a fixed estimate in documentation.

If your AI credit balance is exhausted mid-run, the loop halts with a clear blocked-state message and you can top up before continuing.

Stop conditions in detail

A turn is considered to have produced evidence if at least one of these happened:

• A backtest result was returned (run_backtest, evaluate_strategy_candidate, verify_strategy_candidate)
• A strategy candidate was committed (commit_candidate)
• A research lesson was saved (save_research_lesson)
• The Case Study mission state advanced (a research-plan step completed, mission baseline was set, etc.)

The Case Study ledger advances when any of those produces a durable record (a new CaseStudyVersion, a new mission-state row, etc.).

If both checks fail — a turn produced only inspection-only tool calls and no ledger change — the loop stops to avoid burning tokens on no-op iterations. This is a feature, not a bug; if you want the agent to keep going, type continue and it will plan a different approach on the next turn.

Limits on what the agent will do on its own

The agent will:

• Run backtests, evaluations, verifications, and analyses
• Propose strategy variants and commit them as CaseStudyVersion rows
• Stage verified candidates into your editor for your review
• Save research lessons documenting dead-ends
• Update the Case Study mission state

The agent will not, without your explicit action:

• Save a candidate as the active editor tab’s permanent code (you accept the staged candidate)
• Send orders to your broker (the lab does not execute live orders)
• Modify your TradingView account
• Charge you for anything beyond your configured AI credit budget
• Continue past its stop conditions
• Use a different Case Study than the one you’ve selected

Related

Engine vs TradingView Why the native engine and TradingView can produce different results

Algo Lab Overview The full lab — case studies, variants, optimisation, market data

Algo Lab Quick Start Run your first backtest before invoking the agent

Risk disclaimers The legal warnings covering AI-assisted code and research use