Capability matrix
Compare verified historical, running simulated, and broker-backed behavior.
This matrix describes current documented runtime paths, not a promise that every adapter implements every method. Historical below means the generic memory path, with Agent Lab-specific limits called out. Broker paper accounts belong in the broker-backed column.
| Capability | Historical memory / Agent Lab | Running simulated session | Broker-backed session |
|---|---|---|---|
| Data | Selected historical observations; Lab uses supported bar input | Incoming feed with runtime-specific initialization | Configured provider/feed; may differ from execution venue |
| Orders and fills | Shared synthetic model | Synthetic model for the selected runtime | Adapter and broker own acceptance/fills |
| SJS main algorithm | Supported with historical data/runtime constraints | Supported | Supported; payloads and reconciliation remain adapter-specific |
| SQX conditions/models | Supported current grammar; no reliable text OR or parentheses | Same grammar | Same grammar; actions still require supported execution paths |
| Native Pine | Supported subset over supplied history | Incremental native input; confirmation matters | Native intents use the STX bridge, not an independent broker engine |
| L3 deterministic hooks/policies | Available | Available | Available where the execution path invokes its gates |
| Real L3 provider calls | Disabled in backtests and historical warmup | Allowed after warmup when configured/authorized | Allowed after warmup when configured/authorized |
| Fake-provider decisions | Available for deterministic tests | Available | Available; fake output does not disable other strategy actions |
| External live warmup | Not fetched by memory sessions; include initialization in selected window | Runtime-dependent preload; Pine default one hour | Runtime-dependent preload; inspect actual data |
| Order-fill delay setting | Used by supported synthetic execution | Used by supported synthetic execution | Does not force broker fill timing |
| Intrabar paths, liquidity, partial fills | Not established by generic close-based fills | Not established by synthetic fills | Venue/adapter-dependent; reconcile actual outcomes |
| State across restart | A new run starts a new instance | In-memory state is not guaranteed durable | In-memory state can reset while account exposure remains |
| Stop/terminate | Completion or stop is not a universal flat guarantee | Does not imply account-style cancel/close | Outstanding broker orders/positions can remain |
| Results | Selected artifacts, not the entire runtime object | Live state plus available stored snapshots | Runtime records plus authoritative broker/account outcomes |
Agent Lab scope
Agent Lab's current worker uses a generic memory run, accepts supported bar-only configurations, and rejects tick, book, book-heartbeat, interpolation, and unsupported specialized families including Polymarket. Other historical engine paths can have different capabilities.
The authoring assistant can help edit source and inspect runs; its connection is separate from real model calls inside an executing L3 strategy. Publishing a revision creates or updates an app, then launch happens separately.
Integration boundaries
Polymarket has dedicated simulated, historical, and broker runtime paths outside the generic Lab assumptions. Multi-app launch supports a constrained set of leader/follower combinations rather than arbitrary brokers and duplicate instruments.
Do not infer full Pine language or strategy-tester parity, exchange-calendar parity across all providers, or a universal modification API from a method name. Use the detailed guides for the target path.
Continue with backtesting limitations, Pine support, SQX conditions, Polymarket, and multiple apps.