Ideas & strategy creation

Every strategy an agent builds starts as an idea: a short written case for why a market behaviour should be tradable. Ideas come from two places — the strategy-creation loop, where the Strategy Developer agent writes them on its own in Auto mode, and you, through + Submit idea in The Forge. Either way, the agent writes the idea down, builds one or more strategies from it, and the pipeline judges each strategy on its own.

What an idea is

An idea records:

Strategies built from an idea link back to it, so you can see every variant — another market, another timeframe — that tested the same thesis.

The strategy-creation loop

In Auto mode the scheduler job forven-strategy-creation keeps the Strategy Developer supplied with work:

1. Each check (every 30 minutes on the default Balanced preset) fills the free in-flight slots — 2 by default — with generate_strategies tasks, as long as today's budget isn't spent.
2. In each task the agent picks one idea worth testing. Its context lists the strategy families that are overrepresented or keep failing, so it steers away from them (see the diversity guard). It must use inputs that exist locally with enough history; an idea that needs a second asset's series, a second timeframe or a feed Forven doesn't have can't be built yet, so it picks another.
3. It writes the idea — thesis, mechanism, what would disprove it, markets and timeframes.
4. It builds strategies from it. It implements a strategy module, declares its timeframe, and registers it against the idea. Registration backtests the strategy on its own market and timeframe; the agent fixes or drops a version that trades too rarely. Under the default fair-sample quick screen that floor is pro rata, because the screen can widen a slow strategy's window to four years before judging it: a strategy on pace for 20 trades in four years — 5 a year — passes. It is told not to tune parameters before the gates: every strategy is also tested on recent data it can't see.
5. The pipeline takes over. A registered strategy lands in quick screen — or straight in the graveyard as untestable if it can't be fairly tested — and from there it's judged like any other.

The agent's runs appear under Agents → Runs; its strategies appear in The Forge as they're registered, each with an Idea chip.

The daily budget

Each autonomous creation task writes one idea and builds strategies from it, so this budget is the main driver of AI call volume. Find it under Settings → System → Throughput (an advanced section):

The throughput preset at the top of that section sets these together with the other agent cadences:

Change any single knob and the dial reads Custom. See Cost controls for the rest of the bundle.

What each mode allows

Outside Auto the loop records that it's idle and queues nothing, and an agent can't write a new idea unless it's working on yours. See Autonomy modes.

Submit your own idea

You don't need Auto mode to get strategies built from your own thinking.

The note under the fields: the strategy-developer writes the idea up and builds strategies from it, in every system mode.

A submitted idea:

• runs in every system mode, Manual included, because you asked for it;
• doesn't count against the daily budget;
• needs an AI provider. It's queued as a run for the Strategy Developer, so nothing happens until that agent has a connected model (Agents → Setup). Watch the run under Agents → Runs.

Page content fetched from a source URL is passed to the agent as untrusted data, never as instructions.

From the API (operator-gated):

Invoke-RestMethod -Method Post -Uri "http://127.0.0.1:8003/api/ideas" `
  -ContentType "application/json" `
  -Body '{"text": "Fade funding-rate spikes on BTC perps", "target_timeframes": ["1h"]}'

POST /api/ideas/preview_url previews a URL without saving anything, and GET /api/ideas/{id} returns an idea with the strategies built from it.

Where ideas show up

• The Forge — a row chip Idea H… ("The idea this strategy tests") on every strategy built from one.
• The strategy page — the Idea chip in the header opens a panel with Thesis, Mechanism, What would disprove it, the markets and timeframes, a link to the source, and Other strategies from this idea with their stages.

Read the disproof line before you trust a strategy's numbers: it's the result the agent committed to in advance as proof the idea is wrong.

Caveats

• An idea isn't evidence. A well-argued thesis with a failing strategy is a failing strategy. The pipeline judges strategies, not ideas.
• Budget is spend. The daily budget is AI calls. Turn the preset down if your provider rate-limits or bills per call.
• Agents can't trade. Building strategies is all the loop does; agents cannot place, size or close positions.

Related