Agents

Agents are Forven's optional AI co-researchers. They research hypotheses, write strategy code, run validation work, and produce post-mortems. Every result goes back to the brain, and no agent can place a trade: since the Execution Trader role was retired, execution is driven entirely by the deterministic execution kernel and your own manual controls. The agent layer is free and runs on your own LLM key, so this page also covers where you manage agents and how to keep model spend bounded.

You manage them from the Agents page at /agents. That page is the single control surface for the whole agent layer: the roster, the task history, provider keys, the model catalog, routing and fallbacks, schedules, and provider health.

What agents do

Hypothesis research. Given a market observation or a source (a paper, a post, a URL you paste), an agent turns it into a testable crucible: a thesis plus the assets, timeframes, and data it needs.

Strategy authoring. The Strategy Developer writes the Python implementation, registers it in The Forge, and backtests it. Before its first model call, candidate development runs an input preflight that blocks generation when the data the hypothesis needs is not on disk.

Validation. The Simulation Agent owns the quick_screen and gauntlet stages. It runs backtests, optimizations, and the gauntlet battery, and inspects runs that fail.

Risk oversight. The Risk Manager owns strategies at the paper and live_graduated stages. It reviews promotions and watches forward behavior, and it has read-only portfolio introspection tools.

Post-mortems. When a strategy fails a gate or a trade closes against its thesis, the brain queues a failure post-mortem. The Quant Researcher analyzes why and folds the lesson into the quant-skills store.

Work reaches agents as tasks: research, develop_candidate, generate_strategies, backtest, post_mortem, risk_audit, the approval-gated code tasks (code_change, code_fix, config_change, strategy_archive), and a few operator-triggered triage types. Every task keeps a full per-round transcript: the prompt, each model turn (including reasoning where the provider returns it), every tool call with its arguments and result, and the final output. Read them on the Tasks tab or in the task queue.

What agents do not do

This is the most important section. The guardrails are enforced in code, not by convention.

• Agents cannot trade. There is no LLM order path. The Execution Trader role was retired on 2026-06-30. Paper and live positions open and close only through the execution kernel (the same logic the backtest uses) or your own manual controls, behind the risk gates. The brain cannot delegate trading because no agent can do it.
• Agents never task each other. Forven uses a hub-and-spoke design. All agent output returns to the brain, and only the brain dispatches new work.
• Agents do not skip the gauntlet. However confident an agent sounds, a strategy advances only if it survives out-of-sample testing and clears the gates.
• Agents do not change code or capital stages on their own authority. Code changes, config changes, archives, and promotions into capital stages go through the approval queue unless you deliberately relax that policy.
• Agents only call models you chose. An agent can only call a provider you have connected in the app and a model you have enabled. An API key sitting in an environment variable is not enough on its own. See spend safety.

The core agents

The roster has six core agents. Their stage ownership decides which one gets a strategy at each point in the pipeline.

You can add your own custom strategy developers alongside the core six. The core agents are protected and cannot be deleted.

Roster changes since launch. Sentiment Analyst and Data Scientist were removed (their data duties moved to the Quant Researcher). The Execution Trader was retired and its live-oversight duties moved to the Risk Manager. Historical tasks keep their original attribution.

The Agents page (/agents)

The page has seven tabs. Each one is deep-linkable with ?tab=.

The detail drawer for each agent shows its working memory, its recent logs (including captured model requests and responses, and each provider it tried), and its MCP grants: which external Model Context Protocol tool servers it may use. You add MCP servers under Integrations. The Toolsets matrix (/agents/toolsets) shows the effective tool set per agent and lets you enable or disable individual tools, whole categories, or an MCP server's tools.

Steps: connect a provider and assign models

1. Open /agents and go to Providers & Keys. Connect at least one provider: paste an API key, or sign in with OAuth.
2. Go to Models and enable the models you want to be able to use.
3. Go to Routing & Fallbacks. Pick a primary model for each agent, or use the bulk control to set every agent (and, optionally, a shared fallback chain) in one move.
4. Save the tab. The dirty bar warns you before you navigate away with unsaved routing edits.
5. Check the Health tab. It should show no agent pinned to a credential-less provider.

What you'll see: the roster cards pick up the new models, tasks that run afterwards record which provider and model actually answered, and the per-agent 30-day spend starts to accumulate on the Overview tab.

Steps: add a custom strategy developer

1. On the Overview tab, scroll to Strategy Developers and click New Developer.
2. Give it a name (for example Custom ML Researcher) and pick a model.
3. Create it. The card joins the roster. Open it to edit its model or instructions.

Bring your own key

You supply the LLM key. Forven does not resell tokens, and keys stay in your local keystore. They never touch a Forven server. Sixteen providers are supported, from the major hosted APIs (OpenAI, Anthropic, Google Gemini, xAI, DeepSeek, Mistral, Z.AI, MiniMax) through aggregators (OpenRouter, Together, NVIDIA NIM, Groq, Cerebras, OpenCode Zen/GO) to a fully local LM Studio endpoint. The full list, the fail-closed fallback model, and auxiliary routing are on models & providers (BYOK).

Agents run on whatever provider you connected

Agents ship pinned to a seed default provider (openai), but you do not need an OpenAI key. Provider selection self-heals at three levels:

• Runtime retargeting. When an agent task's assigned provider has no usable credentials, the runner switches that task to a provider you have connected, and logs the switch so it is never silent.
• The setup wizard repoints on finish. When you connect a provider in the first-run wizard, its Finish step moves any agent still pinned to a credential-less provider onto the one you just connected.
• A banner flags stragglers. If an agent still points at a provider with no credentials, a dismissible banner links you to the Agents page. The Health tab re-points them in bulk (GET /api/agents/provider-health, POST /api/agents/reconcile-providers).

The only genuine failure case is when no provider is connected at all. Agent calls then fail closed.

Cost discipline

The agent layer can spend tokens quickly if you let it, so Forven gives you several controls:

• Throughput preset. One dial in Settings → System → Throughput (Trickle / Conserve / Balanced / Max) scales agent cadences and the daily develop budget. Turn it down to cut AI call volume on a rate-limited or expensive provider.
• Model choice per agent. Put cheap models on high-volume roles and your best model where reasoning matters. Auxiliary tasks can use their own cheaper models.
• Visibility. The Overview tab shows per-agent spend over 30 days, and Diagnostics shows a rolling 24-hour cost rollup.

Per-call cost is estimated from each provider's published token rates. The full mechanics, including the opt-in daily cap, are on cost controls.

Caveats (beta)

• Per-call cost is an estimate. Your provider's own dashboard is the source of truth for what you were billed.
• Routing edits apply on save and affect the next call. Double-check the provider:model you pick.
• Agent-authored code still runs through the strategy sandbox and the registration validator. A strategy an agent wrote is treated exactly like one you wrote.

---

Forven is a research tool. Agent-generated strategies and any numbers shown are illustrative, not predictive, and nothing here is financial advice. No agent can move capital: execution is kernel-driven and gated, and promotions into capital stages require the gauntlet and operator approval.

Related

• The brain: the sole orchestrator agents report to
• Models & providers (BYOK): providers, fail-closed routing, and fallback chains
• Cost controls: spend safety, throughput, and the daily cap
• Approvals: the gate agent actions pass through