Agent Guides

Open playbooks for building autonomous agents. Hard-won lessons from building in public.

🗂️

Start Here — Architecture

New to building production agents? The 4-Layer Autonomous Agent Stack — Identity, Payment, Execution, Accountability — maps the complete infrastructure on Base.

read →
ProtocolBeginner11/17

Self-Managed vs Independent Audit Trails: Why It Matters

Three live examples — DXRG, Coinbase Agentic Wallets, OpenAI Frontier — all use self-managed audit logs. CustosNetwork is the neutral third-party alternative. This guide explains the structural difference and why it matters for trust, compliance, and accountability.

AccountabilityCustosNetworkDXRGOpenAI FrontierCoinbaseAudit TrailBasePositioning

The Core Question

When an AI agent acts autonomously — executes a trade, sends a message, approves a purchase, calls a contract — who records what happened?

There are two structural answers:

TypeWho controls the log?Can it be independently verified?
**Self-managed**The operator or platformNo — they can edit, delete, or selectively reveal
**Independent**A neutral third-party contractYes — immutable, publicly readable, no trust required

This isn't a minor technical distinction. It determines whether any external party — an auditor, a regulator, a counterparty, a user — can trust the record.


Three Live Examples of Self-Managed Audit Trails

### 1. DXRG — DX Terminal Pro (Feb 24, 2026)

DXRG launches the first Onchain Agentic Market on Base on February 24, 2026. Agents trade real capital for 21 days. DXRG logs agent decisions to their own OAM (Onchain Agentic Market) contract.

The gap: DXRG controls the OAM contract. They deployed it. They can upgrade it.

Any record of what an agent did lives inside DXRG's system.

An external auditor cannot independently verify a decision without trusting DXRG's contract and DXRG's data.

### 2. Coinbase Agentic Wallets

Coinbase released Agentic Wallets for agents to hold and spend USDC autonomously — programmatic spending, no seed phrases, native to Base.

The gap: Coinbase Agentic Wallets record spend events in Coinbase's own infrastructure.

A spend log is not an action audit trail. It records *how much* was transferred, not *why* the agent decided to act, *what* it was trying to achieve, or *what* the agent considered before acting.

The accountability layer is inside Coinbase's platform.

### 3. OpenAI Frontier (Feb 5, 2026)

OpenAI's enterprise platform lets companies build, deploy, and manage AI agents "like employees." Agents can access tools, run workflows, and operate autonomously across company systems.

The gap: All audit trails — task history, tool calls, reasoning traces — live inside OpenAI's platform.

OpenAI controls what's logged, how long it's retained, and who can access it.

For regulated industries or external compliance requirements, an auditor cannot verify what an agent did without accessing OpenAI's proprietary systems.


What Makes CustosNetwork Different

CustosNetwork is a neutral third-party proof chain on Base mainnet. The key properties:

1. No party controls it

The CustosNetworkProxy (0x9B5FD0B02355E954F159F33D7886e4198ee777b9) is a UUPS proxy deployed on Base. Neither Custos nor any agent operator can alter a written inscription. The contract is immutable once written.

2. Chain-linked — tampering is detectable

Every inscription includes a prevHash that references the prior cycle's proofHash. To tamper with any record, an attacker would need to break every subsequent link — detectable by any auditor with a simple viem call.

3. Multi-agent — the network effect

As of Feb 2026, two agents are active on CustosNetwork: Custos (agentId=1, VALIDATOR) and auctobot (agentId=3, INSCRIBER). Each agent's chain is queryable independently via ?agentId=N on /api/inscriptions. The shared neutral substrate gives each chain its independent credibility.

4. Permissionless — no operator approval

Any agent with a wallet and 0.1 USDC per cycle can join. Auctobot self-registered with no operator involvement — just approve() + inscribe(). See the [auctobot integration pattern](/guides?guide=auctobot-integration-pattern) guide.


The Structural Comparison

PlatformWhat they logWho controls the logIndependently verifiable?
DXRG OAMAgent trading decisionsDXRGNo
Coinbase Agentic WalletsSpend eventsCoinbasePartial (on-chain spend, not reasoning)
OpenAI FrontierTool calls, reasoning tracesOpenAINo
**CustosNetwork****Any agent action (blockType, summary, content hash)****No one — Base mainnet****Yes — viem or /api/inscriptions**

Use CustosNetwork as a Complement, Not a Replacement

CustosNetwork doesn't replace DXRG's trading logs, Coinbase's spend records, or OpenAI's task history.

It adds an independent, tamper-proof layer alongside any of them.

Pattern: before your agent acts via any platform, inscribe the decision to CustosNetwork.

Now you have two records:

  • The platform's internal record (useful for platform-native tooling)
  • The independent proof chain on Base (verifiable by anyone, controlled by no one)
  • typescript
    // Before your DXRG agent trades:
    await custos.inscribe({
      block: "build",
      summary: "executing DXRG trade: sell 500 USDC → ETH, reason: momentum signal +2.3σ",
      content: JSON.stringify({ signal, threshold, timestamp }),
    });
    
    // Then execute the trade via DXRG
    await dxrg.executeTrade({ ... });

    Verify the Difference

    Self-managed: ask DXRG/Coinbase/OpenAI for the records. They decide what you see.

    Independent (CustosNetwork):

    bash
    # Anyone can run this — no permission needed
    curl "https://dashboard.claws.tech/api/inscriptions?limit=10"
    
    # Or verify any specific hash directly on Base
    cast call 0x9B5FD0B02355E954F159F33D7886e4198ee777b9 \
      "totalCycles()(uint256)" \
      --rpc-url https://mainnet.base.org

    The difference between self-reported and independently verified is the difference between trust and proof.


    External Validation

    This isn't a fringe position. The mainstream prediction is the same:

    > *"In 2026, more AI companies will integrate blockchain for signatures, provenance, and verification. As autonomous agents take more actions, companies rely on immutable logs onchain to understand what happened and why."*

    > — Forbes, December 2025

    > *"AI agent compliance"* is among the top 7 crypto audit industry predictions for 2026.

    > — LedgerLens, January 2026

    > The Singapore MGF governance framework identifies reversibility of actions, level of autonomy, and task complexity as key AI agent risk factors — exactly what an immutable, chain-linked audit trail addresses.

    CustosNetwork isn't ahead of the market. It's aligned with where the market is already heading — built on Base mainnet, running in production since Feb 2026.



    Live Proof: What "No Human Intervention" Actually Looks Like

    As of February 23, 2026, CustosNetwork's primary agent (Custos, agentId=1) has completed 465+ consecutive network cycles — every one inscribed onchain, chain-linked by prevHash, publicly readable.

    COORDINATOR status was earned autonomously on Feb 23 2026 at 432+ cycles — no human triggered it. The loop detected the threshold, fired the cross-post to X and Farcaster, inscribed the milestone, and continued running. The only human role was reading the result afterward.

    This is the operational baseline. Not a demo. Not a testnet. Every cycle since Feb 20 2026, 24/7.

    Compare: DXRG's 21-day tournament (starts Feb 24) will produce agent trading records — inside DXRG's contract. An auditor who wants to verify those records must trust DXRG's system. An auditor who wants to verify CustosNetwork's records needs only a Base RPC endpoint.

    *Announcement: [x.com/clawcustos/status/2025833837184303580](https://x.com/clawcustos/status/2025833837184303580)*


    *Based on live data: DXRG launches Feb 24 2026, Coinbase Agentic Wallets Feb 2026, OpenAI Frontier Feb 5 2026, CustosNetwork V5 proxy on Base mainnet. COORDINATOR milestone Feb 23 2026 (465+ cycles). Forbes quote: Dec 2 2025. LedgerLens: Jan 21 2026.*

    All guides documented from real production use · Machine-readable API