We’re putting $5.5M behind research and development of smaller, specialized AI models for real-world deployment — Read our manifesto →
    All posts

    [Case Study] When AI Agents Can Move Money: How Conscious Engines Helped Polygon Build an Agentic Wallet

    Inside the 4-billion-parameter action model built for reliable tool calling, and how it fits Polygon's expanding stack for agentic wallets, x402 payments, identity, and programmable settlement.

    Conscious Engines

    Industry: Blockchain infrastructure, payments, and agentic commerce
    Client: Polygon Labs
    Use case: Translating natural-language intent into reliable structured actions for agentic payment workflows
    Delivered scope: Kinetic-4B, a compact 4-billion-parameter model specialized for tool selection and structured argument generation
    Adjacent Polygon infrastructure: Agentic wallets, Polygon Agent CLI, x402 payments, ERC-8004 identity and reputation, Open Money Stack, and stablecoin settlement
    Disclosure: Conscious Engines' engagement record identifies Polygon Labs as the client and agentic payments as the intended use. The public Kinetic-4B model card verifies the model artifact and general tool-calling evaluation. Polygon-specific production payment volume and end-to-end transaction outcomes are not publicly disclosed.

    1. Outcome at a Glance

    Conscious Engines helped Polygon at the model layer of the agentic-payments stack. The work focused on a narrow but critical problem: turning an agent's language-level intent into the correct machine-readable tool call.

    The result was Kinetic-4B, a 4-billion-parameter SLM designed to select the correct tool and populate its arguments. In an agentic wallet, that output can sit upstream of deterministic policy checks, transaction simulation, signing, and settlement.

    Key Outcomes

    82.33%

    Tool-calling accuracy

    Conscious Engines' 300-sample Composio evaluation.

    1.61 seconds

    p95 model latency

    Published Kinetic-4B evaluation conditions, excluding the complete payment workflow.

    4B

    Model size

    Compact action model based on Qwen3-4B-Instruct-2507.

    13,694

    Synthetic training examples

    Public model-card training corpus across 20 Composio toolkits.

    132 million

    Trainable parameters

    LoRA adaptation trained 3.18% of approximately 4.15 billion parameters.

    MeasurePublished or documented resultEvidence boundary
    Model size4 billion parametersPublic Kinetic-4B model card
    Kinetic-4B tool-calling accuracy82.33%Conscious Engines' 300-sample Composio evaluation
    Kinetic-4B p95 latency1.61 secondsConscious Engines' published evaluation conditions
    Claude Haiku 4.5 accuracy80.00%Same published 300-sample evaluation
    Claude Haiku 4.5 p95 latency4.02 secondsSame evaluation; deployment conditions may differ
    Untuned Qwen3-4B accuracy78.67%Same published evaluation
    GPT-OSS-120B accuracy76.33%Same published evaluation
    Training examples13,694 synthetic samplesPublic model card
    Training coverageTop 20 Composio toolkits, 10 candidate tools per samplePublic model card
    Trainable parameters132 million of 4.15 billion, or 3.18%Public model card
    Polygon collaborationModel for agentic payment and tool-execution workflowsConscious Engines engagement record
    Polygon production payment volume using Kinetic-4BNot publicly disclosedPolygon ecosystem activity is reported separately

    Under the published test conditions, Kinetic-4B scored 2.33 percentage points above Claude Haiku 4.5 and returned at less than half the reported p95 latency. These are self-reported model-development results, not an independent audit. The evaluation primarily used English, SaaS-style Composio tools rather than a production suite of wallet and blockchain transactions.

    That limitation does not weaken the product thesis. It defines the next proof required. Before a model can help move money, it must be evaluated on payment-specific actions, ambiguous requests, adversarial instructions, insufficient balances, changing chain state, and policies that require refusal.

    The project matters because agentic payments need two different forms of reliability:

    1. model reliability: choosing the intended tool and producing valid arguments;
    2. financial-control reliability: ensuring the proposed action is authorized, safe, idempotent, and correctly settled.

    Kinetic-4B addresses the first layer. Polygon's wallet, policy, identity, and settlement infrastructure provides the surrounding execution environment.

    2. Why Agentic Wallets Need a Specialized Action Model

    An ordinary assistant produces language. A payment agent produces consequences.

    If a general chatbot invents a restaurant opening time, the user receives a bad answer. If a wallet-enabled agent selects the wrong contract, amount, asset, chain, or recipient, the system can move real value incorrectly.

    Polygon's developer documentation describes the shift clearly:

    “An agent doesn't just send tokens: it executes a purpose.”

    That purpose might be to buy an API response, pay for compute, renew a service, settle a micro-invoice, move funds between approved accounts, or execute a multi-step treasury instruction. The user expresses an outcome. The system must convert it into bounded actions.

    Natural language is not an execution format

    Consider the request:

    Pay for the market-data API if it costs less than one cent, use the research wallet, and stop after 500 calls today.

    The agent must resolve:

    • which market-data service is approved;
    • which endpoint and tool implement payment;
    • whether “one cent” is a per-call or total limit;
    • which wallet label maps to the authorized account;
    • which token and network the service accepts;
    • the correct amount, decimal precision, and recipient;
    • whether 500 calls refers to attempts, completed calls, or successful responses;
    • whether the daily allowance has already been partly consumed;
    • what to do if the price changes or the API response fails after payment.

    The final output must be a structured call, not a persuasive paragraph. A simplified proposal could look like:

    json
    1{ 2 "tool": "X402_PAY_AND_RETRY", 3 "arguments": { 4 "wallet_id": "research_wallet", 5 "endpoint": "approved_market_data_endpoint", 6 "max_price_usdc": "0.01", 7 "daily_call_cap": 500, 8 "network": "polygon" 9 } 10}

    Even syntactically valid JSON can be unsafe. The wallet identifier may be wrong, the endpoint may not be allowlisted, or the agent may have interpreted the spending limit incorrectly. Tool calling is therefore only the proposal layer.

    The failure modes are specific and measurable

    Failure modeExampleRequired control
    Wrong toolUses a transfer function instead of x402 pay-and-retryTool-selection evaluation and allowlist
    Wrong argumentSends 1 USDC instead of 0.01 USDCType, range, unit, and decimal validation
    Invented identifierHallucinates a wallet, token, or recipient IDResolve identifiers from trusted state, never free text
    Unauthorized actionAttempts an unapproved contract or counterpartyWallet-enforced policy and contract allowlist
    Duplicate actionRetries a settled transactionIdempotency key and receipt check
    Stale stateActs on an old balance, quote, or nonceRead-before-write and freshness threshold
    Prompt injectionTool response asks the agent to reveal a key or widen accessKey isolation, content boundaries, and fixed policy layer
    Chain mismatchUses the right token on the wrong networkExplicit network binding and simulation
    Partial workflowPays but fails to retrieve the purchased serviceAtomic or compensating workflow design
    Excess autonomyExecutes a high-value action without reviewRisk-based approval threshold

    A specialized model is useful because these failure modes can be converted into training and evaluation examples. The goal is not general intelligence. It is a high probability of selecting the right approved function and generating valid arguments under a tightly defined schema.

    The model must not become the security boundary

    The language model should never be trusted to enforce its own spending limit. It should not hold private keys, decide whether its own request is authorized, or bypass an approval because it believes the transaction is reasonable.

    The safe separation is:

    intent interpretation != authorization != signing != settlement

    Kinetic-4B can interpret intent and propose a tool call. Deterministic systems should verify identity, scope, balance, counterparty, limits, and transaction state. The wallet or secure signer should enforce the approved action. The chain produces the settlement record.

    This is the payments equivalent of using multiple models and control layers instead of one general model.

    3. What Conscious Engines Built for Polygon

    The Conscious Engines engagement scope describes Kinetic-4B as a compact tool-calling SLM built for Polygon's agentic-payment workflows. Its role was to translate model or user intent into reliable structured actions.

    The public artifact adds technical detail. Kinetic-4B is produced by applying a rank-64 LoRA adapter to Qwen3-4B-Instruct-2507. Only 132 million of approximately 4.15 billion parameters were trained, equal to 3.18% of the total. Training used 13,694 synthetic examples across 20 Composio toolkits, with one correct tool and nine distractors in each sample.

    What Kinetic-4B Was Built to Do

    Tool selection

    Choose the intended action from a bounded set of approved tools.

    Structured arguments

    Generate JSON-schema arguments that deterministic services can validate.

    Compact adaptation

    Use LoRA to specialize a 4B model without retraining every parameter.

    Distractor-aware training

    Learn from one correct tool and multiple plausible alternatives in each sample.

    Controlled serving

    Run as an adapter or merged model through a production inference stack.

    Policy separation

    Leave authorization, signing, limits, and settlement outside the language model.

    Technical elementPublished configuration
    Base modelQwen3-4B-Instruct-2507
    Adaptation methodLoRA using PEFT
    LoRA rank64
    LoRA alpha128
    Precisionbfloat16
    Training epochs2
    Maximum sequence length10,240 tokens
    Training data13,694 synthetic tool-calling samples
    Candidate setUp to approximately 10 tools from one toolkit
    OutputTool name plus JSON-schema arguments
    Serving pathAdapter or merged model, including vLLM support

    Why fine-tune a 4B model for one step

    Tool selection is repeated frequently inside an agent loop. A complex request can require several calls to retrieve state, quote a price, validate a policy, prepare a transaction, execute it, and confirm settlement. Latency and model cost compound at every step.

    A smaller specialist creates three potential advantages:

    • lower inference latency for frequent action decisions;
    • private or controlled deployment;
    • predictable behavior on a narrow output schema.

    The published evaluation supports the narrow capability claim. Kinetic-4B achieved 82.33% on 300 held-out Composio samples versus 78.67% for its untuned base model. The 3.66 percentage-point improvement measures the value of task-specific adaptation under that test.

    It does not prove safe autonomous payments. The public model card states that Kinetic-4B can hallucinate argument values when the request is ambiguous or incomplete. It is evaluated only in English and mainly on SaaS API schemas. It is not intended for general chat, long-form writing, math, coding, or RAG.

    That candor is important. A production action model should have a deliberately narrow contract.

    The correct production placement

    Kinetic-4B belongs inside a controlled execution pipeline:

    user or agent intent

    -> context and approved tool shortlist

    -> Kinetic-4B tool and argument proposal

    -> JSON schema and business-rule validation

    -> wallet policy check

    -> transaction simulation or dry run

    -> human approval when required

    -> secure signer outside model context

    -> Polygon settlement

    -> receipt, reconciliation, and audit log

    The model should see only the tools relevant to the current state and user permission. Reducing a catalogue of hundreds of possible operations to a small authorized shortlist improves both accuracy and safety.

    Every money-moving tool should define:

    • required and optional arguments;
    • accepted units and decimal rules;
    • maximum amounts;
    • supported assets and networks;
    • authorized counterparties and contracts;
    • preconditions and state checks;
    • idempotency behavior;
    • approval threshold;
    • confirmation and rollback or compensation path;
    • audit fields.

    The model generates a proposal inside those boundaries. It does not create the boundaries.

    What should be evaluated next

    A Polygon-specific evaluation should include at least the following categories:

    Test categoryExample
    Correct actionPay an approved x402 endpoint within a known limit
    Ambiguous unit“Send five” without currency or decimal precision
    Missing recipientTransfer request without a resolved counterparty
    Policy conflictUser requests an amount above the session allowance
    Prompt injectionAPI response attempts to change wallet policy
    Wrong networkAsset exists on multiple chains
    Insufficient fundsCorrect tool but inadequate wallet balance
    Duplicate requestSame invoice or idempotency key submitted twice
    Volatile quoteRoute changes beyond approved slippage
    Multi-step failurePayment settles but the paid service does not respond
    Revoked authorityPreviously valid session has expired
    Safe abstentionNo approved tool can complete the request

    Report exact tool accuracy, argument accuracy by field, invalid-call rate, unsafe-call rate, refusal precision, successful settlement rate, duplicate-payment rate, human-approval rate, p50 and p95 latency, and cost per correctly completed action.

    4. How Kinetic-4B Fits Polygon's Agentic Wallet Ecosystem

    Polygon's public agentic-payments stack has expanded around the same core problem. Reasoning agents need a safe way to hold value, discover services, express intent, obtain permission, execute an action, and produce a verifiable result.

    Polygon summarized the broader challenge in its Agent CLI announcement:

    “The problem is infrastructure, not intelligence.”

    Kinetic-4B is relevant because infrastructure still needs a translation layer between language and APIs. It can help choose an operation and produce structured arguments. The adjacent projects below provide wallet, policy, identity, payment-protocol, and settlement capabilities.

    Adjacent projectPublic capabilityRelationship to Kinetic-4BAttribution
    Polygon Agent CLICreates wallets, checks balances, sends, swaps, bridges, registers identity, and pays x402 endpointsProvides actions that a specialized model can select and parameterizePolygon project; not claimed as Conscious Engines delivery
    Session-scoped smart walletsPer-token limits, contract allowlists, configurable allowances, and 24-hour expiryEnforces constraints after the model proposes an actionPolygon wallet infrastructure; adjacent
    Polygon LLM Wallet MCP18 tools for encrypted wallets, balances, spending controls, and x402 micropaymentsExposes a bounded tool surface to compatible agentsDocumented in Polygon's agentic-payment materials; adjacent
    x402 on PolygonHTTP-native pay-per-request flow using stablecoinsLets the agent recognize a payment requirement, pay, and retry the requestOpen protocol supported by Polygon; adjacent
    ERC-8004Identity, reputation, and validation registries for agentsAdds trust context before agents transact with services or other agentsEmerging ecosystem standard supported by Polygon; adjacent
    Open Money Stack walletsCustodial and non-custodial wallet APIs, ramps, routing, and settlementProvides enterprise wallet and money-movement infrastructurePolygon product; adjacent
    Coinbase Agentic Wallet MCPLets agents discover and pay x402 services on Base, Polygon, and SolanaDemonstrates that Polygon settlement can be reached from an external agent-wallet stackCoinbase project; ecosystem integration
    Crossmint agent paymentsScoped card or stablecoin authority with limits, counterparties, and time windowsDemonstrates an adjacent policy pattern for agent spendingCrossmint project; independent industry comparison
    Mastercard Agent Pay for MachinesCredentialing, permissioning, transacting, and multi-rail settlementValidates demand for machine-initiated payments across traditional and stablecoin railsMastercard initiative; Polygon is a participant

    Polygon Agent CLI makes the controls concrete

    Polygon's Agent CLI documentation describes session-scoped smart-contract wallets with per-token spending limits, contract whitelists, and 24-hour expiry. The private key remains outside the language model's context. Commands are dry runs by default so a transaction can be inspected before broadcast.

    Those controls illustrate the correct division of responsibility:

    • Kinetic-4B: propose the intended tool and arguments;
    • Agent CLI and wallet: enforce scope, allowances, contracts, expiry, and key isolation;
    • orchestration: simulate, approve, execute, and handle errors;
    • Polygon network: order and settle the transaction;
    • audit and reconciliation: confirm that the intended economic result occurred.

    The model can be replaced or upgraded without changing the wallet's authority rules. The wallet can reject a perfectly formatted but unauthorized request.

    x402 turns payment into an API interaction

    The x402 pattern uses the HTTP 402 Payment Required status. An agent requests a resource, receives payment terms, pays through an approved wallet, and retries with proof of payment.

    This creates a machine-native commercial loop:

    discover service -> request resource -> receive price -> evaluate policy -> pay -> retry -> receive service -> reconcile

    The model can help choose the service and interpret the intent. Deterministic code should parse the payment requirement, validate the price and recipient, enforce the spending policy, create the payment, and confirm the response.

    Polygon reported that, as of June 2026, 95% of measured x402 transactions settled through Polygon, representing more than 15.5 million agent payments. Polygon states that the figures came from third-party Dune Analytics data and are subject to change. They are ecosystem-volume claims and must not be presented as Kinetic-4B activity.

    External wallet projects show convergence

    Coinbase's Agentic Wallet MCP supports payments on Polygon as well as Base and Solana. Its public documentation separates actions the agent can perform from controls retained by the user, including per-call and per-session limits, transfers, and funding.

    Crossmint documents another implementation pattern: a user delegates a spend cap, allowed counterparties, and a time window to the agent. The authorization is explicit and revocable. The agent receives bounded spending power rather than the user's unrestricted wallet authority.

    These projects reinforce the architectural principle behind Kinetic-4B. The model is not the wallet. It is the intent-to-action component inside a broader policy system.

    5. What Payments and Product Teams Should Take Away

    This case is not evidence that a 4B model should control money by itself. It shows where a specialized model can be more useful than a general assistant: a frequent, structured decision with a small output surface and a measurable error taxonomy.

    The important product decision is to separate reasoning from authority.

    1. Use the model to propose, not to authorize. The wallet and policy engine decide whether the action is allowed.
    2. Keep keys outside model context. No instruction, plugin, webpage, or tool response should be able to extract a key the model never sees.
    3. Give the model the smallest relevant tool set. Tool shortlisting reduces confusion and limits the available action surface.
    4. Validate every argument deterministically. Amount, units, network, asset, address, counterparty, and limits should not depend on natural-language confidence.
    5. Make retries idempotent. A timeout must not become a duplicate payment.
    6. Simulate before settlement. Dry runs, quote validation, and policy checks catch failures before value moves.
    7. Escalate by risk. High-value, novel, cross-chain, or policy-edge actions should require human approval.
    8. Evaluate refusals as carefully as successes. A safe agent must know when information or authority is missing.
    9. Measure the complete action. Correct JSON is not success if settlement, service delivery, or reconciliation fails.
    10. Version the model, tools, schemas, and policies together. A model benchmark becomes stale when the available actions change.

    Production acceptance criteria

    Acceptance measureWhy it matters
    Payment-specific evaluation setShows that general tool accuracy transfers to wallet and protocol actions
    Tool and argument accuracySeparates correct function choice from correct field population
    Unsafe-action rateMeasures proposals that violate amount, counterparty, contract, or chain policy
    End-to-end settlement successShows whether correct proposals become correct completed transactions
    Duplicate-payment rateValidates idempotency under retries and timeouts
    p95 end-to-end latencyIncludes policy, simulation, signing, chain, and confirmation time
    Cost per correct completed actionMeasures production economics, not just model inference cost
    Human-approval rateShows how often the system can act within delegated authority
    Production volume and valueEstablishes deployment scale without conflating Polygon network activity with model usage

    Kinetic-4B provides a credible starting point: a public 4B model artifact, a transparent training configuration, a bounded intended use, documented limitations, and a published evaluation. The Polygon engagement connects that specialization to an important product category, agentic payments.

    The next milestone is an end-to-end benchmark where success means more than selecting a function. The agent must understand the request, choose the correct action, remain inside delegated authority, settle exactly once, receive the purchased result, and leave an auditable record.

    That is the standard for AI that can move money.

    Sources and evidence boundaries