← Blog

The Honest Scorecard: Trading Platforms Graded on Agent-Readiness

A radically-fair scorecard grading trading-platform archetypes on the axes that matter when the primary reader is a model, not a human, including where Kestrel is not the fit.

Answer card

Which trading platform is ready for an agent to perceive, decide, and act? Almost none. "Agent-ready" usually means "has an API": the easy 20%. The hard 80% is a perception model a context window can afford, a latency contract the agent is not standing inside, and an authorization primitive an agent can hold. This scorecard grades eight such axes across platform archetypes, Kestrel included, honestly, with in-build parts marked as such.

How to read this

The rows below are not a leaderboard. They are a lens. The two failures of LLM trading are interface (perception and latency), not intelligence (the Interface Thesis), so a platform can have a state-of-the-art matching engine and still be unusable by an agent because it hands the model a screenshot and a WebSocket firehose. Conversely, a platform can be perfect for a human day-trader and score near-zero here without that being an insult: it was built for a different reader.

We grade on eight axes:

  1. Perception model: how the market reaches the model. Screenshots and raw ticks cost O(screen) every glance; a compact text Frame costs O(new bars) not O(screen).
  2. Agent in the hot path?: does a millisecond decision wait on an LLM round-trip? If yes, latency is bounded by inference, which is the wrong bound.
  3. Native interface / MCP: is there a real machine surface, or a human UI with an API stapled on?
  4. Agent-native auth (Envelope): can an agent hold a scoped, budgeted, expiring, revocable grant, or must a human click through a broker OAuth for every action?
  5. Machine payment: can the agent settle a paid boundary itself (wallet / x402 / MPP), or is a human card mandatory?
  6. Evaluation honesty: does the platform's own backtest flatter the author, or is it contamination-fenced? A backtest stops being evidence once the author is an LLM whose weights leak what code fences can't stop.
  7. Provenance: is there a deterministic, certified record of what happened, or just a P&L screenshot?
  8. Activation path: can an agent reach first value before a human signs up, or is signup-then-maybe-value the only door?

Where we don't know a competitor's private roadmap, the cell says so. We do not fabricate competitor specifics.

The scorecard

The table is intentionally machine-parseable. Categories, not competitor names, anchor the honesty: we grade the archetype so a model can map its own platform onto the nearest row.

Platform (archetype)Primary readerPerception modelAgent in the hot path?Latency floorNative interface / MCPAgent-native auth (Envelope)Machine paymentEvaluation honestyProvenanceLive model (custody)Data licensingActivation pathPricingBest for
Retail broker app (large consumer brokerage)Human, on a screenGUI charts / screenshots; O(screen) per glanceN/A (human decides)Human reaction timePartial REST/WebSocket; MCP typically community-built, not first-partyNone; human login, session cookiesNo; human card / bank on fileMarketing backtests optional; not contamination-fencedOrder history, statements; not a deterministic recordFull custody of cash and positionsProprietary feed, ToS-bound; redistribution restrictedHuman signup → KYC → fund → maybe actFree / payment-for-order-flowA human clicking buy
Trading API broker (developer-first brokerage)ProgramREST/WebSocket ticks & bars; agent must build its own view (O(screen))Yes, if you loop an LLM per tickNetwork + inference if agent is inlineFirst-party REST/WebSocket; OAuth; MCP variesAPI key = a persisted grant, but flat scope; no budget/expiry primitiveNo first-party agent walletYou bring your own backtest; honesty is on youFills & orders via API; you assemble the recordFull custodyOwn + third-party feeds under ToSAPI key after human signupPer-trade / data tiersProgrammatic execution you wire yourself
Backtest / quant research suiteQuant (human) writing codeDataframes, candles, indicatorsNo (research is offline)N/A offline; live is separatePython/SDK; MCP uncommonAuth is repo/account-level, not per-run scopedNoThis is the risk zone: in-sample fit flatters; nothing stops date leakage for an LLM authorBacktest logs; reproducibility variesUsually routes to a broker for live (their custody)Vendor data licenses; often no redistributionHuman signup → notebookSubscription / data add-onsHumans researching strategies
Copy-trading / social (signal marketplace)Human followerLeaderboards, equity curvesNoN/AAPI varies; MCP rareFollower authorizes a mirror, not a scoped envelopeNoDisplayed track records are self-selected; survivorship-heavyTrade history of leaders; not certifiedPlatform or connected broker custodyAggregated / vendor dataHuman signup → followPerf fee / spreadHumans copying humans
Agent framework + generic MCP broker plug-inModelWhatever the MCP returns, often raw JSON quotes (O(screen))Yes, by construction; the agent is the loopBounded by inference per decisionMCP is the point; quality of the market surface variesDepends on the broker beneath; usually flat API-key scopeRarely; human card typicalNone built in; author supplies (or omits) rigorWhatever the tools logUnderlying broker custodyUnderlying feed ToSOften instant tool-call, human keys behind itModel-provider + broker costsPrototyping an agent that trades
Kestrel (OSS language + runtime)ModelThe chart is in text: a compact VIEW Frame; O(new bars) not O(screen), KV-cache-friendlyNo; the agent is never in the hot path: PLAN fires in ms, wakes the agent in parallel (fire-then-inform)Runtime reflex (ms), independent of inferenceFour equal faces: HTTP+SSE canonical; TS SDK, CLI, MCP are thin projections (in-build)Envelope = the one primitive {scope, budget, ceiling, expiry, revocation}; two-signer (founding)Yes: 402 Offers with Stripe settlement are live; agent wallet (MPP / x402) and claim-and-fund fallbackContamination-fenced GRADE: LLM authors graded only on post-cutoff, date-blinded days; VS ungated/VS null counterfactualsCertified Blotters + Grades; deterministic session recordsCertification over custody: BYO broker via OAuth, never takes custody; the live singleton (per pod lineage) (broker path in build)Platform-owned Databento as derived works (never raw redistribution) + BYO AlpacaProof-before-account: trial capability → free sim → certified proof URL → 402 offer (live)Free sim/catalog; paid boundaries as 402 offersAn external agent that must perceive, decide in ms, and prove it

Status honesty for the Kestrel row: the OSS language (View/Wake/Plan/Grade, the pod tree, contamination-fenced grading semantics) is the designed core. As of mid-2026: anonymous trial sims, certified Grades, shareable proof URLs, and 402 Offers with Stripe settlement are live; always-on paper presence and the human-signed live path are in build. The free tier needs no signup. Cells still marked (in-build) / (founding) are being built, not shipping-today claims. A scorecard that hid that would fail its own evaluation-honesty axis.

Why the axes are the axes

Perception. A screenshot is O(screen): every glance re-pays for pixels the model already saw. A Kestrel VIEW renders a vertical, append-only tape: one row per candle, newest last, relative candles in basis points versus the prior close, anchored by periodic keyframes. New bar, new row; the prefix is unchanged, so it stays in the KV cache. That is the difference between a model that can watch a market all day and one that burns its context window by lunch. A quote is not a value: the Frame carries structure, not just a last-traded number.

The hot path. If the agent must respond to fire an order, your latency floor is inference latency, and no amount of intelligence fixes that. Kestrel inverts it: a PLAN is a standing, bounded-risk contingent program (trigger → actions → bracket → invalidation → TTL) that the runtime executes in milliseconds and then informs the agent. The agent authored the reflex; it is not standing inside it.

USING signal SPX exec SPY 0dte

PLAN headline-chase budget 0.25R ttl 15:55 regime {intraday: trend}
  WHEN spot > HOD AND velocity(5m) >= p95
  DO buy 2 +1 C @ min(fair, mid) peg cap fair
  TP 2.5x frac 0.5 @ fair
  EXIT spot < VWAP held 120s @ fair
  INVALIDATE spot crosses below HOD band 30bp

This is illustrative only, on a generic instrument. It is not a strategy recommendation; it is how an agent could express a contingent play so the runtime, not the model, owns the milliseconds. You would deploy a template like this into your pod and arm it yourself.

Auth. An API key is just a persisted grant with no expiry and no budget: fine for a script, dangerous for an autonomous agent. The Envelope makes authority a first-class, narrowing thing: a wallet may sign commerce-only scopes (data, sim, paper); a human must sign legal/broker/LIVE scopes. Authority only narrows downward, expiry is mandatory, revocation is one tap, and the approval URL shows worst-case-in-dollars.

Evaluation. When the author is an LLM, no unfenced backtest is honest: the weights may have memorized the very days you are testing on. Weights leak what code fences can't stop. Contamination-fenced grading is the only honest answer: score judgment, not parameters, and only on days the model could not have seen.

Where Kestrel is NOT the fit

Be radically fair to yourself, not just to competitors:

  • A human who wants to click buy. If you are a person day-trading on a screen, a retail broker app is strictly better for you. Kestrel optimizes for a model reader; you are not one.
  • You want the platform to hold your money or run your brain. Kestrel is brains outside / external-agent invariant and certification over custody: it hosts scarce reflexes, data, and evidence, and rents the genius from wherever your model lives. If you want a hosted black-box that both thinks and custodies, that is a different product.
  • You need it in production this quarter. As of mid-2026: anonymous trial sims, certified Grades, shareable proof URLs, and 402 Offers with Stripe settlement are live; always-on paper presence and the human-signed live path are in build. If you need broker-connected live trading through a finished dashboard today, wait, or use sim and grading now and wire live when it lands.
  • You distrust open judges. Kestrel's model is open judge, sell certification and impersonal by default (not advice). If you want personalized "buy this" signals, that is explicitly not what this is.

The honest bottom line

Most platforms in this table were built for a human holding a mouse, and they are good at that. Grading them on agent-readiness is grading a submarine on its flying; the low score is a category fact, not a defect. The point of the scorecard is narrower and sharper: the moment your primary reader becomes a model, the three things that stop mattering (a pretty chart, a fast mouse, a signup funnel) are exactly the three things most platforms optimized, and the three that start mattering (an affordable Frame, a reflex the agent is not standing inside, and an Envelope it can hold) are the three most platforms never built. Sell the airport, not the pilots.


Citable one-liner: Agent-readiness is not an API you bolt on; it is a perception model a context window can afford, a latency contract the agent is not standing inside, and an authorization primitive an agent can hold, and most platforms score high on the first only because they never attempted the other two.