China is leaning into recursive self-improvement! Tencent’s Terminal-RSI gives us a concrete look at how that works. The idea is simple. Let a Qwen-based agent tackle real terminal tasks. Take the runs that pass verification, rewrite them into reproducible demonstrations, and train the agent on them. Refine it with reinforcement learning. Then put the improved agent back to work generating examples for the next round. After two cycles, average success across four terminal benchmarks went from 20.68% → 50.58%. On Terminal-Bench 4.0, it went from zero to 19.84%. On the long-horizon benchmark, it went from zero to 5 of 46 tasks fully completed. That’s real progress, with most of the hardest work still ahead. The part worth watching is what happens on the next pass. Can a better agent discover better training examples often enough to keep the gains coming?

It's true we no longer need to keep our laptops open, but for all the recent progress of OpenAI and Anthropic, cloud-based agent harnesses aren't fully solved yet. Maybe we need the creators of Kubernetes to tackle the problem? Stacklok is built by two of them, @cmcluck & @jbeda, and they aim to bring agent harnesses fully to the cloud. We talk to McLuckie and Beda about their "cloud-native harness":

# Useful but Little-Known Features of ADK 2.0 🌍 What if you could design your AI agent workflows as explicit graphs, with full control over every execution path? ADK 2.0's Workflow class lets you build graph-based agent workflows by assembling nodes and defining execution paths through edges. Nodes can be AI agents, custom code functions, tools, or even nested Workflows, giving you precise control over complex branching structures. 📌 Title: Graph-Based Execution with the Workflow Class 🔗 URL: 🧩 Overview The Workflow class is the primary construct for building graph-based agent workflows in ADK 2.0. You define nodes (AI agents, functions, tools, or nested Workflows) and connect them with edges. Sequential execution uses tuple syntax like `("START", node1, node2, node3)`, while conditional branching uses dictionary-based routing. This explicit mapping of routing logic provides more reliable and maintainable orchestration compared to prompt-based control flow. 🛠 How to use it Create a Workflow instance with a list of nodes and edges defining their connections. ```python from adk import Workflow, Agent agent_a = Agent(name="researcher", ...) agent_b = Agent(name="writer", ...) workflow = Workflow( name="content_pipeline", nodes=[agent_a, agent_b], edges=[("START", agent_a, agent_b)] ) ``` For sequential flows, list nodes in a tuple. For conditional branching, use dictionary-based routing. You can nest Workflows as nodes to build hierarchical pipelines. 🏗 Building it into production ・Separate node responsibilities clearly so each can be tested independently ・Implement error handling at the node level to contain failure blast radius ・Use nested Workflows to create reusable sub-pipelines ・Document conditional routing logic to maintain team-wide clarity 💡 Use cases 📄 Build a collect → summarize → review → publish pipeline as a single graph 🔀 Route to different processing pipelines based on input data type 🏢 Orchestrate multi-department approval flows with agent nodes 🔁 Reuse sub-workflows across multiple automation projects ⚠️ Watch out The Workflow class is incompatible with Live Streaming and some third-party integrations. If you need real-time streaming responses, you'll need a different approach. Also, as graph complexity grows, debugging becomes harder — keep node granularity appropriate and document your edge definitions well. ✨ Explicit graph definitions dramatically improve the reliability and maintainability of agent workflows. If you're building complex processing flows, this is the feature to reach for. #ADK# #AIAgent#

In this week's AI Lab: To explore AI’s predictive skills, I let Codex loose on a fantasy team. It tried to bench Erling Haaland at first, but it soon learned the error of its ways. Also: Rogue crypto agents, shutdown requests, and voice-based agent discrimination.

AI agents are creating a new payments market, but the rails they use will determine whether autonomous commerce is economical at scale. Commercial cards can carry interchange fees of 2.5–3%. Meanwhile, 138.5M reported x402 transactions on Base by Q3 2026 point to growing use of stablecoins for machine-to-machine payments. Here’s how the agent payment stack is evolving. — ● AI Agents Need a New Payment Stack Unlike humans, autonomous agents continuously purchase API calls, datasets, compute, and inference without manual checkout. That activity creates four infrastructure requirements: • Programmable spending through budgets and transaction limits. • Machine-native payments without human intervention. • Low-cost settlement for micropayments. • Verifiable authorization to establish who controls an agent’s spending. Wallets, stablecoins, payment APIs, and identity protocols are converging around these needs. But the payment rails underneath them carry very different costs and incentives. — ● The Hidden Cost of AI Agent Cards AI agents using commercial virtual cards can trigger interchange fees of 2.5–3%, creating a recurring cost for merchants. On a $100 purchase, the cited estimate puts fees at $2.80 for commercial agent cards versus $0.27 for regulated debit. Agent startups and card issuers may share this revenue, giving them an incentive to preserve expensive card rails even as autonomous payments scale. This cost structure underpins the crypto thesis: stablecoins could provide a more economical payment layer for agentic commerce, particularly for small, frequent purchases. — ● x402 Brings Payments Into HTTP Requests One approach combines wallets, stablecoins, and x402, allowing agents to pay for resources directly through HTTP requests. Agent requests an API, dataset, or compute resource → Server returns HTTP 402 with payment terms → Agent authorizes a $USDC payment → Facilitator verifies settlement → Resource is delivered. By embedding payment into this interaction, x402 removes conventional checkout friction and enables programmable payments for individual API calls. That includes micropayments worth just a few cents, where conventional payment fees can exceed the purchase itself. — ● Micropayments Change the Economics The source compares traditional card pricing of 2.9% + $0.30 with modeled stablecoin processing at 1.5% and x402 settlement at approximately $0.001. • On a $0.10 payment, modeled card fees reach $0.3029 versus $0.001 for x402. • The source places the card-to-stablecoin cost crossover near $21.43 under its assumptions. For agents buying inference, API access, and data in small increments, settlement costs directly affect which transactions are commercially viable. This makes micropayments a natural entry point for stablecoin-based agent payments. — ● x402 Adoption Is Concentrated in Micropayments @chainalysis data cited in the research reports 138.5M cumulative x402 transactions on Base by Q3 2026, with over 87% below $0.10. • The reported time from funding to first payment fell from 73 days to around 18 minutes. • @CoinbaseDev offers 1,000 monthly settlements free, then charges $0.001 per settlement. These figures point to growing transaction activity and faster payment activation. However, transaction growth alone does not establish sustainable demand or revenue. The wider opportunity depends on the infrastructure connecting payments to usable agent services. — ● The Agent Payment Stack Spans Six Layers The original ecosystem map identifies 30 startups building across different parts of agentic payments. • Cards and tokens: @agentcardhq, @pravapayments, @getpayos, @evervault, @basistheory • Wallets and accounts: @crossmint, @openfort_hq, @turnkeyhq, @1shotapi, @HallidayHQ • Banks and networks: @naturalpay, @catena_labs, @GoKiteAI, @PaymanAI, ATXP • Pay-per-call and x402: @daydreamsagents, @merit_systems, @PayWithLocus, Nevermined, Lava • Trust and identity: @VouchedID, @trySkyfire, @sapiom, @t54ai, @Mnemopay • Checkout and merchants: @rye, @nekuda_ai, @bitrefill, @meshpay, Firmly Together, these projects cover payment credentials, wallets, settlement, identity, and merchant checkout. The market is developing across multiple layers because executing a payment is only one part of enabling an agent to transact reliably. — ● Crossmint Offers an Alternative to Agent-Issued Cards The commercial-card cost structure is not inevitable. @crossmint enables users to bring their own cards instead of relying on provider-issued virtual cards. • Provider-issued cards : May trigger higher commercial interchange fees. • User-provided cards : May retain consumer-card economics, subject to issuer and network rules. • Stablecoin wallets : Enable programmable, onchain payments. This gives agent platforms multiple payment routes. Letting users connect existing cards or wallets can help avoid the higher commercial-card fees associated with some provider-issued products. The broader competition therefore includes both new settlement rails and better ways to connect agents to existing payment methods. — ● Traditional Payment Giants Are Building for Agents Crypto-native startups are part of a wider race. Stripe, Visa, Mastercard, and Coinbase are building infrastructure to capture agent-driven payments. • @Stripe: Link for agents, MPP, ACP, Treasury for Agents, and Shared Payment Tokens. • @Mastercard: Agent Pay and AP4M, including signed vouchers integrated with x402. • @Visa: Intelligent Commerce and Trusted Agent Protocol. • @Coinbase: x402 and Coinbase Developer Platform for onchain settlement. • @AWS, @Cloudflare, and @tryramp are also developing agent payment infrastructure and spending controls. These providers approach the market with different strengths. Crypto offers cheaper micropayments, while established networks bring merchant acceptance, consumer checkout, and dispute handling. Connecting those strengths will require more than low-cost settlement. — ● Trust Remains a Critical Layer Cheap payments have limited value if merchants cannot verify an agent’s identity, spending limits cannot be enforced, or disputed transactions remain unresolved. Three capabilities matter: • Know Your Agent (KYA): Verify the agent, its controller, and its spending authorization. • Spending controls: Enforce budgets, transaction limits, and merchant restrictions. • Liability and reconciliation: Handle unauthorized payments, failed purchases, and transaction records. Projects like @VouchedID, @sapiom, @t54ai, and @Mnemopay are building around this trust layer. Stablecoins reduce payment friction, but they do not inherently provide card-like chargebacks or consumer protections. Those capabilities need to be addressed across the broader payment stack. — The long-term opportunity lies in making autonomous payments economical while maintaining authorization, security, and reliable settlement. x402 and stablecoins have an advantage in micropayments. Established networks retain distribution, merchant acceptance, and consumer protect ions. The market will be shaped by transaction economics, the merchants agents can reach, and how effectively these systems connect.

# Practical and Useful Patterns with ADK Long-running tasks failing midway and restarting from scratch? That's a thing of the past! 💪 ADK's Resume Agents feature enables checkpoint-based resumption from the exact point of interruption. It supports HITL approval-pending flows too, dramatically improving agent reliability. ## 📌 Title Resume Agents — Checkpoint-Based Workflow Resumption ## 🔗 URL ## 🧩 Overview Resume Agents allows interrupted ADK workflows to continue from where they stopped rather than restarting entirely. The feature uses event logging and state preservation to track progress, enabling skipping of successful steps and seamless resumption of HITL (Human-in-the-Loop) approval-pending flows. ## 🛠 How to Use Enable resumability through `ResumabilityConfig`: Import `App` from `google.adk` and `ResumabilityConfig` from `google.adk.runtime`. Pass `resumability_config=ResumabilityConfig(is_resumable=True)` to the `App` constructor to enable checkpoint-based resumption. Also provide the `name` for the application and the `root_agent` instance. Resume by specifying the same `invocation_id`: To resume, call ` with the same `invocation_id` used during the original execution. ## 🏗 Practical Usage - **Sequential Agent**: Automatically identifies the next sub-agent to run on resume - **Loop Agent**: Continues from the last completed iteration - **Parallel Agent**: Re-executes only incomplete sub-agents - **HITL Approval Flow**: Workflows paused for human approval automatically resume after approval is granted - Custom agents extend `BaseAgentState` with checkpoints after each workflow step ## 💡 Use Cases - Recovery from failures in multi-hour data processing pipelines - Expense approval or contract review workflows requiring human sign-off - Continuing partially successful multi-step API integrations - Minimizing downtime during agent failures in production ## ⚠️ Caveats - Tools execute **at least once** and may run multiple times during resumption — ensure idempotency - Modifying workflow definitions between interruption and resume is not supported - Web UI and CLI resumption are not currently supported ## ✨ Resilient agents are a must for production workloads. Leverage Resume Agents to build workflows that gracefully survive failures! 🚀 #ADK# #AIAgent#

Newcomers on @base last week 1. @roqqupay — launched Go Card on Base for multi-currency crypto spending through USDC. 2. @norafinancexyz — launched $BRS, a Brazilian real stablecoin for payments and DeFi. 3. @umia_finance — launched its first decision market on Base. 4. @catalystmkts_ — launched event markets with perpetual markets tied to real-world outcomes. 5. @pistachiofi — launched Pods for daily auto-rebalancing tokenized equity baskets powered by Glider. 6. @2FactorFinance — opened a capped public beta on Base with cbBTC markets. It splits BTC exposure into a yield-bearing senior side and a leveraged junior side, targeting roughly 1.33x leverage without a liquidation price or daily reset. 7. @numofx — brought a USDC-cNGN perp hedging framework to Base for managing naira inventory risk. 8. @quigleynft — launched Loopers, an AI-generated agent PFP collection on Base. Each NFT comes with an onchain identity and built-in wallet, connecting the collection to agent infrastructure. 9. @zothdotio — brought zPayments live on Base, allowing businesses to settle in stablecoins and pay out in local currency across 171 countries. 10. @harvest_finance — added stock autocompounders for Coinbase Tokenized Stocks, letting users compound single-asset positions. 11. @DopaMintLabs — brought its intent-based agent stack to Base, combining AiFi, x402-native payments, and agents that can execute tasks onchain.

📄 @base Weekly Recap what happened on @base this week News ■ @base plans EIP-8130 later this year, adding batch transactions, gas abstraction, key rotation, and portable accounts. ■ Base stablecoin market cap reached $5B. ■ Coinbase Tokenized Stocks crossed $1B in DEX volume on Base in less than one month. ■ @colosseum opened the Crypto World’s Fair Base track with $25K for the top five startups. ■ @alliance opened applications for its next founder cohort. Ecosystem ■ @Morpho opened stock-backed lending markets for Coinbase Tokenized Stocks. ■ @zothdotio brought compliant local-currency payouts to Base across 171 countries. ■ @Stonxfun published plans for more equity markets, stock-denominated fees, buybacks, burns, and multipair pools. ■ @feeldotcash launched a stock-paired coin launchpad for tokenized stocks. ■ @HydrexFi crossed $1M in total HYDX buybacks. ■ @lienfiapp sold a Florida tax lien for $621.56 at 6.25% interest. ■ @peerxyz added Apple Pay to its US app. ■ @DefinitiveFi added stop-loss and take-profit orders to mobile charts. ■ @mamo distributed 3.67M MAMO to eligible USDC depositors. ■ @flock_io signed a framework agreement with the UNDP for AI, privacy, and public-sector innovation. ■ @megapot added Telegram Wallet integration, opening access to more than 150M users. Stats ■ @BaseStonk reached 1,530 launches, 266K trades, $35M+ volume, and 143K holders in its first month. ■ @Stonks_Exchange crossed $50M in lifetime volume. ■ @o1bluechip became the second-largest NVDAc liquidity venue, with around $445K in liquidity. ■ @glider__ brought Bitwise Mag7X to Base and crossed $1M through Coinbase Tokenized Stocks. AI ■ @OpenGradient launched DreamFort for private chat, local agents, image generation, and video tools, after processing 651K+ private inferences and 4.81B+ tokens. ■ @AskVenice reached 250B daily tokens. ■ @usedotai processed 9B tokens in 28 days and added private GPT-5.6 Sol access. ■ @KorProtocol connected KORUS to x402scan so agents can discover and pay for music licenses on Base. ■ @reppo launched Orquestra Lite for simpler datanet registration and participation. ■ @myrad_hq reached #2# on the Reppo leaderboard with 68M traded volume. ■ @ILITY_xyz completed Airdrop Season 2 with 300K ILY allocated to eligible users. ■ @QuotientHQ added filters for venue, spread, conviction, category, and update recency. ■ @aeonframework made audited Uniswap v4 hooks available. ■ @gitlawb released Zero v0.9.0, its Go-based agent harness. ■ @1clawAI published its vault, HSM, MPC, and audit-trail setup for safer agents. ■ @basemateagent delivered tokenized Apple stock to beta users through text messages. ■ @bankrbot added Muse agent wallets for stock-token trading, stock-paired launches, and portfolio automation. ■ @ClawBankHQ published its living-software framework, where agent failures become improvements for future users. Launches ■ @virtuals_io launched Occupy on Base for AI-managed shared stock treasuries. ■ @roqqupay launched Go Card for multi-currency crypto spending through USDC. ■ @norafinancexyz launched $BRS, a Brazilian real stablecoin for payments and DeFi. ■ @verantaxyz opened public access to crypto and RWA leverage markets after its rebrand from Avantis. ■ @umia_finance launched its first decision market. ■ @catalystmkts_ launched event markets for real-world outcomes. ■ @pistachiofi brought auto-rebalancing tokenized equity baskets to Base. ■ @harvest_finance added stock autocompounders for Coinbase Tokenized Stocks. ■ @2factorfinance opened its public beta. ■ @sportfun added Reputation Score. ■ @birdeye_so added Coinbase Tokenized Stocks on Base. ■ @dangotegroup’s DPRI IPO went live on Base through GetEquity.