The chat interface is great for building and validating a strategy. But once a strategy is live, you need something that watches the market continuously, adapts when conditions change, and acts when signals fire, without you being glued to a screen. That's what ambient agents do.

An ambient agent is an autonomous quantitative researcher and trader that operates during Indian market hours (and off-hours for research). It wakes up when something relevant happens, decides whether to act, and goes back to sleep. Unlike a static algorithm, it reasons about market conditions before every decision.

How ambient agents differ from the chat interface

Chat interfaceAmbient agent
You drive every iterationAgent wakes autonomously on events
Runs one strategy at a timeManages a portfolio of paper deployments
Stateless across sessionsBeliefs and findings persist across wakes
You interpret resultsAgent posts findings and acts on them
No capital allocationFixed capital pool per agent

The chat interface is where you research, backtest, and validate. Ambient agents are where you put validated strategies to work under real market conditions (paper money, no real capital at risk).

What an ambient agent does during a typical day

  1. Day start wake: every morning before market open the agent reviews open deployments, re-anchors on the current regime (volatility, trend direction), and arms the day's alerts and interest watches.
  2. Event-driven wakes: the agent wakes only when something it armed fires: a price alert, a pattern condition, a deployment trade, or a market-regime transition. If nothing fires, it doesn't wake (and doesn't burn budget).
  3. Decisive action or stand-down: on each wake the agent looks at the context, checks its beliefs and the opportunity board, and decides: deploy, monitor, adjust, or stand down. A wake with no action is normal and expected.
  4. End of day: the agent closes any intraday positions before market close, records learnings in its missed-alpha ledger, and arms next-day watches.

Launching an ambient agent

From the chat interface, once you've backtested a strategy you're happy with, ask:

"Launch an ambient agent to trade EMA crossover strategies on large-cap NSE stocks. Allocate ₹2,00,000 in paper capital."

The platform generates an agent configuration with:

  • Mission: a plain-language description of the strategy mandate. The agent reads this on every wake to stay on-task.
  • Capital allocation: the total paper money pool. The agent enforces this strictly; it will not deploy more than the allocation.
  • Watchlist: instruments to monitor. The agent can expand this dynamically via discover_market_universe if your mission is broad.

You can also launch an agent from the Ambient tab in the sidebar, where you'll see all active agents, their status, and their aggregate P&L.

The mission contract

Every agent operates under a mission contract that governs its behaviour:

  • Max active deployments: how many live paper positions it can hold simultaneously.
  • Max daily deployments: a daily cap on new positions opened.
  • Quality gate: minimum backtest quality tier (A/B/C/D) for a strategy to be deployed. The agent won't deploy a strategy that hasn't been backtested and doesn't meet the threshold.

The agent reads its contract on every wake. If a deployment would breach the contract, it refuses and records the rejection in the missed-alpha ledger.

The token budget

Each agent runs under a $10/day token budget (USD, per IST calendar day). This covers all LLM calls the agent makes during the day: research, analysis, and deployment decisions. When the budget is reached, the runtime stops calling the model; any deployed paper strategies keep running on their own. The budget resets at midnight IST.

This makes the system economical: a quiet day with no signals might consume $0.50. An active day with multiple deployments and monitoring passes might hit $3-5.

Persistent beliefs

Beliefs are the agent's durable memory. When the agent forms a view (for example, "RELIANCE RELIANCE logoNSE:RELIANCE is in an uptrend with strong OI accumulation"), it records this as a belief. On the next wake, the belief is still there. The agent can update, confirm, or retract beliefs as market conditions evolve.

This is what makes ambient agents feel like a real researcher rather than a stateless script: the agent remembers what it decided yesterday, why it passed on a trade, and what it's watching.

Monitoring your agents

The Ambient tab shows all agents with:

  • Status: active, paused, budget_exceeded, retired
  • Capital used vs allocated
  • Total P&L across all paper deployments
  • Activity feed of recent wakes and decisions

Click into an agent to see the full wake history, each decision with its reasoning, and the current opportunity board.

Next steps