5bix Weekly

AI Agents

Autonomous Agents Move From Demo to Production as Industry Finds Its Guardrails

State machines, budget caps and human-in-the-loop checkpoints are replacing the naive ReAct loop. The era of shipping agents that survive real-world traffic has begun.

Every week a new agent framework appears on GitHub. The README shows a screen recording of a chatbot that books flights, writes code and files taxes, all from a single prompt. It is impressive until you try to run it on anything that matters. The flight it books is for the wrong date. The code it writes has a subtle off-by-one. The tax filing hallucinates a deduction.

At 5bix we have shipped agent systems for on-chain monitoring, document processing and customer support in Web3. None of them were easy. All of them required patterns that the framework authors rarely talk about: explicit state machines, retriable tool calls, human-in-the-loop checkpoints and budget caps that prevent a runaway model from burning through an API key in minutes.

The single most important architectural decision for every agent is how it manages its own context. A naive implementation appends every observation to a growing message list until it hits the token limit and crashes. A production agent needs a summarisation strategy: compress older turns, drop tool outputs that are no longer relevant, and maintain a persistent scratchpad of facts that must survive across context window boundaries.

The most reliable agents we have built are not free-form reasoners. They are state machines with a language model at each transition. The model decides which transition to take, but the set of legal transitions is fixed by the developer. This eliminates the failure mode where the model invents a tool that does not exist or skips a mandatory validation step.

An agent that can call an LLM in a loop can burn through hundreds of dollars in minutes if it gets stuck in a retry cycle. Every production agent needs a hard budget cap: a maximum number of LLM calls, a maximum total token spend and a maximum wall-clock time. When any cap is hit, the agent stops, logs its state and alerts a human.

The lesson, after two years of shipping these systems, is that reliability beats autonomy every time. An agent that handles 80% of cases perfectly and escalates the rest is worth far more than one that handles 95% of cases and silently fails on the other 5%. Trust is earned one correct action at a time.

“The loop is trivial. The hard part is knowing when to stop, what to remember and what to forget.”

Smart Contracts

Gas Optimization After Dencun: What Still Matters

EIP-4844 rewrote the cost model for Ethereum. Storage operations still dominate, but the tricks that save gas have changed.

For most of Ethereum’s history, gas optimization was a dark art practised by a small guild of assembly-literate engineers. Much of that folklore is now outdated. The EVM has changed, the compiler has improved and the cost model itself has been rewritten by the Dencun upgrade.

An SSTORE to a cold slot costs 20,000 gas. An SLOAD from a cold slot costs 2,100 gas. By comparison, an ADD costs 3 gas. Storage operations dominate the gas profile of almost every contract, and the most effective optimisation you can do is to reduce the number of them. Struct packing remains valuable — two uint128 values in one 256-bit slot cost one SLOAD instead of two.

Transient storage, introduced in EIP-1153, is the new option for data that only needs to survive within a single transaction. TSTORE and TLOAD cost 100 gas each — a fraction of their persistent counterparts. Reentrancy guards, callback flags and single-transaction accumulators are natural candidates.

“The deployment cost is paid once. Execution gas is paid on every call. Optimise for the thing that repeats.”

On-Chain Data

Substreams Indexes 200M Events Per Hour Without Falling Behind

How a parallelised map-reduce pipeline replaced days-long subgraph syncs with thirty-minute backfills.

Blockchain data is deceptively simple in structure and brutally hard to query at scale. Every block is a list of transactions, every transaction is a list of calls and events. The information is all there — but extracting it in real time, at the throughput of a chain producing a block every two seconds, requires infrastructure that most teams underestimate.

Substreams takes a fundamentally different approach from subgraphs. Instead of processing events one at a time in a single-threaded runtime, it models the indexing pipeline as a directed acyclic graph of map and store modules that can run in parallel across block ranges. The modules are compiled to WebAssembly, which gives them near-native execution speed.

The key insight is that most indexing operations are embarrassingly parallel at the block level. Extracting Transfer events from block 18,000,000 has no dependency on block 17,999,999. A Substreams pipeline can fan out across hundreds of block ranges simultaneously. What took a subgraph three days takes Substreams thirty minutes.

Mobile

React Native’s New Architecture Finally Delivers on Its Promise

JSI, Fabric and TurboModules have matured. The framework that once stuttered on complex animations now matches native performance in production.

React Native has changed more in the past eighteen months than in the previous five years combined. The New Architecture — comprising JSI, Fabric and TurboModules — has moved from experimental flag to default, and with it, the framework’s performance story has been rewritten entirely.

The old bridge serialised every call between JavaScript and native into a JSON message, queued it asynchronously and hoped for the best. JSI replaces this with synchronous C++ bindings that let JavaScript call native methods directly, with no serialisation overhead.

Expo has become the de facto way to build React Native apps. Its managed workflow handles native modules, OTA updates, push notifications and app store submissions with zero Xcode or Android Studio configuration. Over 30% of the top 100 US App Store apps now use React Native in some capacity.

“The framework that once stuttered on complex animations now ships the Instagram and Shopify apps.”

Industry

Week in Numbers

Key metrics from the past two weeks across the ecosystem.

Ethereum daily active addresses crossed 1.2 million for the first time since the 2021 bull run. Base alone accounted for 340,000 of those — more than Arbitrum and Optimism combined. The L2 fee compression predicted by EIP-4844 is working exactly as designed.

Claude 4 extended its context window to 1 million tokens, and Anthropic reports that 38% of enterprise customers are now using it for agent workflows. Meanwhile, OpenAI shipped structured outputs for o3, and Google’s Gemini 2.5 Flash crossed 200 billion tokens per day in inference volume.

In DeFi, total value locked across all chains hit $180 billion, led by Ethereum ($95B), Solana ($28B) and Base ($12B). Uniswap v4 processed its one-billionth swap, and Aave crossed $50 billion in cumulative loans originated.

5bix Dispatch

What We’re Building This Sprint

Updates from the 5bix engineering floor in Mohali.

The agent monitoring dashboard is in internal beta. It tracks state machine transitions, tool call success rates and token spend across all deployed agents. Early numbers show a 3x reduction in debugging time compared to log-grepping.

Our Substreams pipeline for cross-chain ERC-20 indexing hit production this week, covering Ethereum, Arbitrum and Base simultaneously. Backfill completed in 47 minutes — a personal best. Steady-state latency from block production to database write averages 1.8 seconds on Ethereum mainnet.

The Chronicle itself is now a 5bix product. This newspaper-format blog is generated biweekly by Claude, reviewed by the editorial team via Telegram, and published automatically on approval. The entire pipeline — from research to layout to deployment — runs without manual code changes.