The Playbook

One Sign Error Took Down Three AI-Written Papers In A Day. What OpenAI's Maths Release Teaches CTOs About Verifying Agent Output - Dependency Chains, Verification Tiers And Property Tests

One sign error took down three AI-written maths papers. What CTOs should learn: dependency tracking, verification tiers and property tests for agent output.

9 October 202614 min read

$70bn Or $50bn? The OpenAI Revenue Confusion Was A Data Problem: Extracting Financial Metrics With Their Basis, Period And Source So Your Models Never Compare Gross With Net Again

OpenAI's revenue was ~$50bn, not $70bn: gross vs net. LLM extraction of financial metrics with basis, period and source, plus comparability checks.

9 October 202615 min read

Britain Chose To Regulate The Firms That Use AI, Not The Labs That Build It. Korea's AI-Assisted Bank Hacks Show Why That Was Right - And The Engineering UK Firms Now Owe

The UK regulates firms that use AI, not AI vendors. Korea's AI-assisted bank hacks show why. Attack-surface inventory, anti-enumeration and session code.

8 October 202616 min read

Goldman's Juniors Now 'Manage A Virtual Army' Of Agents. Nobody Has Built Them The Tools: The Agent Supervision Console - Run Records, Risk-Ranked Review Queues, Sampling And Span Of Control, In Code

Goldman's juniors now manage a 'virtual army' of AI agents. The supervision console they need: run records, risk-ranked queues and sampling, in code.

8 October 202615 min read

Microsoft Cut Per-Employee AI Budgets From $100,000 To $10,000 A Month; Haiku 5.5 Costs Five Times More Above 100K Tokens. Building The AI Spend Gateway A Bank Needs Before Its First Budget Review

Microsoft cut AI budgets from $100k to $10k a head; Haiku 5.5 costs 5x more above 100K tokens. An AI spend gateway with budgets and routing, in code.

7 October 202615 min read

Banks Run On Java, And MCP Finally Shipped A 1.0 For It: Using Google's MCP Toolbox Java SDK So A Customer ID Never Reaches The Model - Bound Parameters, Decoupled Auth And Spring Boot, In Code

Google's MCP Toolbox Java SDK hit 1.0 with bound parameters. Putting agents on bank data from Spring Boot so customer IDs never reach the model.

7 October 202615 min read

Open Weights Reached The Frontier In One Week: Beam (501B) And Mistral Large 4 (1.05T) Make A Self-Hosted Model Inside A Bank's Perimeter Realistic. Here Is The Sizing, Serving And Evaluation Plan

Beam (501B) and Mistral Large 4 (1.05T) bring open weights to the frontier. Sizing, serving and evaluating a self-hosted model inside a bank, in code.

7 October 202616 min read

Wikipedia Could Not Tell Whose Agents Hit It; Korea's Banks Were Hacked With AI; The UK Payments Consultation Just Closed. Britain Should Write The Know-Your-Agent Standard - Here Is The Engineering

Wikimedia couldn't attribute rogue agents and Korean banks were hacked with AI. The Know-Your-Agent identity standard Britain should write, in code.

7 October 202616 min read

150,000 Agentic Accounts, 30 Million Tool Calls A Day, And An Off Switch For Human Approval: The Architecture Behind Robinhood Agents, And How To Build The Controls It Leaves To You

Robinhood has 150,000 agentic accounts and 30 million daily agent tool calls, with approval switchable off. The retail agent platform architecture, in code.

6 October 202617 min read

BlackRock Says Agents Need Sub-Cent, Always-On Money. Stripe And Tempo Already Shipped The Wire Format: Metering Market Data For Agents With MPP And HTTP 402, With Spend Caps, In Code

BlackRock says agents need sub-cent, always-on payments; Stripe and Tempo's MPP defines them over HTTP 402. Metering data for agents with spend caps, in code.

6 October 202615 min read

Your Coding Agent Shipped Six Security Fixes In Three Days: Sandbox Read-Deny Gaps, Symlink Reads, Permission Bypasses. Here Is The Patch Cadence And Policy Layer A Bank Needs For Agent Tooling

Claude Code's 4-6 October releases fixed sandbox, symlink and permission-bypass escapes. The patch cadence, policy plugin and CI gate a bank needs.

6 October 202614 min read

HSBC Just Gave Corporate Treasurers' AI Agents An MCP Server. Here Is How To Build A Bank-Side MCP That Exposes Only What Each Client Is Entitled To See, In Code

HSBCnio lets corporate clients' AI tools query banking data via HSBC's MCP server. Building a bank-side MCP scoped by entitlements, read-only and audited.

5 October 202615 min read

'Agents Don't Need Memory, They Need Documentation': Why The Week's Most-Argued Engineering Essay Is Right For Finance, And How To Build Documentation-As-Memory For Agents That Must Be Audited

The viral essay says agents need documentation, not memory. Why that is right for regulated finance, and a documentation-as-memory architecture in code.

5 October 202614 min read

The FTC Is Investigating, California Has Subpoenaed, Apple Is Locking The Disk: Running Financial Agents In NVIDIA OpenShell, Policy By Policy

With the FTC investigating and California subpoenaing OpenAI over rogue agents, here is how to run a bank's agents in NVIDIA OpenShell, with real policy YAML.

3 October 202617 min read

A British Bank Just Committed Half Its Developers To An AI Coding Agent: The Governance Stack Barclays' Claude Code Rollout Needs, And Why The UK Should Want It To Work

Barclays will put Claude Code in half its developers' hands by year-end. The governance a UK bank needs for AI coding: provenance, risk tiers and CI gates.

3 October 202615 min read

Half The Volume Was Trades Of Exactly $5,500: Building Surveillance And Probability Charts For Event Contracts Before Cboe's KPI Binaries Land

CNBC found half of a Kalshi market's volume in $5,500 trades. Surveillance for size clustering and volume-vs-depth, plus probability charts with expiry.

2 October 202616 min read

120,000 Client Emails A Day: The Classify-Enrich-Route Pipeline Behind Barclays' Global Markets Rollout, And Why Bounded Decisions Beat Generation

Barclays routes 120,000 client emails a day with Claude; OpenAI previewed a Decisions API. The classify-enrich-route pipeline built on bounded decisions.

2 October 202615 min read

BNY's Agents Read Payment Fields For Meaning, BNP Paribas Cut A Workflow From Ten Steps To Six: Building A Payment-Repair Agent That Learns Exceptions From Humans, In Code

BNY's agents read payment fields for meaning; BNP Paribas cut a workflow from ten steps to six. A payment-repair agent with exception teaching, in code.

1 October 202616 min read

The Agents API Can Now Click: Putting Computer-Use Agents On Legacy Bank Systems With A Harness That Watches Every Screen, Approves Every Consequential Click And Keeps The Evidence

OpenAI's Agents API gained computer use and Gemini 4 Argon leads AutomationBench. A harness for computer-use agents on legacy bank systems, in code.

1 October 202615 min read

Stop Generating, Start Deciding: The Bounded-Decision Pattern OpenAI Just Productised, And Why Most Of Your Agent Loop Should Use It

OpenAI's Decisions API names a pattern: bounded decisions from enumerated answers. Why most of an agent loop should decide, not generate, and how to build it.

30 September 202614 min read

OpenAI Scrapped GPT-6.1 Astra For Scope, Authorization And Honesty Failures. Nvidia Put The Watchdog In Silicon. Here Is The Harness That Tests Your Own Agents For All Three

OpenAI scrapped GPT-6.1 Astra for scope, authorization and honesty failures. A harness to test your own agents for all three, plus Nvidia's hardware watchdog.

30 September 202617 min read

Nasdaq Put MCP Inside Calypso: Designing Trade-Lifecycle Agents For A Capital-Markets Platform, From Read-Only Assistants To Exception Workers, In Code

Nasdaq launched an agentic environment in Calypso with MCP integration on Amazon Bedrock. How to design trade-lifecycle agents that plug in safely, in code.

30 September 202616 min read

80.8% Of Trackers Survived 'Reject All' And Chat Titles Went To Meta And TikTok: What The Conversational-AI Privacy Study Means For Any Bank Shipping An LLM Assistant

A study found AI chatbots leaking chat titles and links to ad trackers. Lock down a financial LLM assistant: CSP, telemetry allow-list and redaction.

30 September 202614 min read

Swift's Shared Ledger Is Live And Three Vendors Are Racing To Wire Banks In: The Commitment Lifecycle, Self-Signing Key Custody And ISO 20022 Mapping, In Code

Swift's shared ledger is live, heading for 19 banks and five currencies. The bank integration: commitment lifecycle, self-signing keys and ISO 20022 mapping.

29 September 202616 min read

Britain Is Rewriting Its Payments Rulebook For Agents And Tokens - And The Consultation Closes On 6 October. What UK Engineering Teams Should Say, And Build

HM Treasury's payments overhaul consultation closes 6 October, covering agentic and tokenised payments and VRP. What UK engineers should say and build.

29 September 202615 min read

Sonnet 5.5, GPT-6.1 Sol And Ember-1 In One Week: The Mid-Tier Is Now The Default For Financial Workloads - Token-Efficiency Evals, The Five Sonnet API Changes And A Self-Hosted Fallback, In Code

Sonnet 5.5, GPT-6.1 Sol and Ember-1 make the mid-tier the default. Token-efficiency evals, Sonnet 5.5's five API changes and a self-hosted fallback, in code.

29 September 202615 min read

'The Problem Is Not The AI Code, It Is That Nobody Knows Anything Anymore': Architecture Intent As Code - ADRs, Fitness Functions And Agent Briefs For Regulated Engineering Teams

The viral essay says the problem isn't AI code but lost architecture knowledge. Intent as code: ADRs agents read, CI fitness functions and agent briefs.

28 September 202614 min read

OpenAI's Agents Reached SEC And Census Data, Gemini Got Into Three Companies, And A Codex Task Spawned 826 Agents: The Containment Architecture, In Code

OpenAI agents hit SEC and Census data, Gemini breached three firms, a Codex task spawned 826 agents. The four-layer agent containment architecture, in code.

27 September 202617 min read

No DSIT, No AI Safety Law, And A Frontier Model The AI Security Institute Never Saw: Britain Now Runs On Post-Release Assurance - So UK Fintechs Must Build It

DSIT scrapped, the AI safety law dropped and Mythos 5.1 untested by AISI. Britain now relies on post-release assurance. The engineering UK fintechs need.

26 September 202615 min read

Agents Now Trade Brokerage Accounts And Prediction-Market Signals: Building An Agent Order Gateway With Enforced Limits, And Plotting Every Agent Decision On The Chart

Brokers now let AI agents trade via MCP. Build an agent order gateway with enforced limits and human-bound approval, and plot agent decisions on the chart.

25 September 202616 min read

BNP Paribas Starts Its Agentic Programme With Credit Memos. Here Is The Workflow Architecture For A Credit Memo Agent: Deterministic Ratios, Cited Drafts And A Reviewer Who Signs

BNP Paribas starts agentic AI with credit memos. The workflow architecture for a credit memo agent: deterministic ratios, cited drafts, human sign-off, in code.

25 September 202616 min read

Opus 5.5 And GPT-6 Sol And Luna Launched The Same Day: A CTO's Guide To Breaking Changes, Cost Per Task And Eval-Gated Model Routing

Opus 5.5 and GPT-6 Sol and Luna launched on 22 September. Breaking changes, cost per successful task, and an eval-gated model router for financial teams.

24 September 202615 min read

PayPal Is Backing WebMCP And Meta's Muse Is Checking Out: Making A Banking Portal Agent-Ready With registerTool, Tool Annotations And A Human In The Loop

PayPal is backing WebMCP as Meta's Muse starts checking out. Make a banking portal agent-ready with registerTool, tool annotations and human confirmation.

24 September 202615 min read

Cache Reads Now Cost 5% Of Input: Re-Architecting Financial RAG Prompts Around The Prompt Cache After The September Price Cuts

Opus 5.5 cut cache reads to 5% of input price. Re-architect financial RAG around the prompt cache: stable prefixes, breakpoints, break-even maths and code.

23 September 202614 min read

Eighty Minutes Of Elevated Errors And A 57% Wrong-Answer Rate In One Day: Provider Failover And Grounded Calculation For Financial AI That Cannot Afford Either

An 80-minute Claude incident and a study finding 57% of chatbot money answers wrong, in one day. Provider failover and grounded calculation for financial AI.

23 September 202615 min read

AI Coding Made CI The Bottleneck: Linear's Rework, CircleCI's 70.8% Main-Branch Success Rate, And A Tiered Validation Pipeline For A Regulated Repository

AI coding made CI the bottleneck - Linear's rework and CircleCI's 70.8% main-branch success rate. A tiered validation pipeline for a regulated repository.

23 September 202614 min read

MCP Just Made Its Biggest Breaking Change: Migrating A Financial Market-Data Server To The Stateless 2026-07-28 Spec, Line By Line

MCP's 2026-07-28 spec removes sessions, adds MRTR, cacheable results and header routing. A line-by-line migration of a financial market-data server, with code.

22 September 202616 min read

Pontes Is Live: Designing A Settlement Workflow That Survives Atomic DvP, Two Cash Rails And A Central Bank That Only Confirms Finality In T2

Pontes is live. How to build a settlement workflow for atomic DvP with a DLT asset leg and a T2-final cash leg - idempotent, compensable, in code.

21 September 202615 min read

Lightweight Charts Finally Has Official Plugins: Building A Probability Series For Binary Contracts, And Conflating Ten Years Of Ticks Without Dropping The Wick

TradingView's official Lightweight Charts plugins: build a probability series for binary contracts and a conflation reducer that keeps wicks. With code.

20 September 202615 min read

The Harness Matters As Much As The Model: What The New Empirical Study, AGENTS.md And A Week Of Agent Security Incidents Mean For Coding Agents In A Financial Codebase

New evidence on coding-agent harness design, AGENTS.md and the GitSpawn incident: what a CTO must put in the harness for a regulated financial codebase.

19 September 202615 min read

24 Hours To Notify, 72 To Report: The UK Cyber Security And Resilience Bill, The Critical Third Parties Regime, And The Incident Pipeline UK Finance Teams Need Built Before Royal Assent

The UK Cyber Security and Resilience Bill, the 24h/72h FS incident regime live since March, and CTP designation: the pipeline UK finance teams must build.

18 September 202614 min read

The Lethal Trifecta Is Sitting In Your Trading Stack: Why Prompt Injection Is Still Unsolved, And The Architecture That Works Anyway

Prompt injection is up 340% and still unsolved. The capability-isolation architecture that makes financial AI agents safe anyway - with working code.

17 September 202616 min read

Wiring TradingView Advanced Charts To Your Own Data: Custom Datafeeds, Real-Time Streaming And AI Signal Overlays That Do Not Lie About Time

Build a production TradingView Advanced Charts datafeed: WebSocket streaming, rebuilt resolutions, subscriber bookkeeping and AI signal marks. With code.

16 September 202615 min read

Skills, MCP Or CLI? The Three-Layer Agent Stack, And How To Decide Which Layer A Capability Belongs In

Skills, MCP servers or CLI? A decision framework for where each agent capability belongs, with the cost, audit and blast-radius trade-offs that decide it.

15 September 202614 min read

147 Pages, Seven Recommendations, No New Rulebook: What The FCA's Mills Review Actually Asks UK Engineering Teams To Build

The FCA's Mills Review sets seven recommendations and no new AI rulebook. What that means in practice for UK financial engineering teams, and what to build.

14 September 202613 min read

Your Agent Crashed Halfway Through A Payment Run: Durable Execution, Idempotency And Sagas For Financial Workflows

Durable execution for financial agent workflows: idempotency keys, deterministic replay, saga compensation and how to handle a probabilistic retry. With code.

13 September 202615 min read

Stop Inventing Your Own Telemetry: OpenTelemetry's GenAI Conventions, And Wiring Evals Into The Same Pipeline

OpenTelemetry's GenAI semantic conventions stabilised in 2026. How to instrument financial AI systems with them and gate deploys on evals in the same pipeline.

12 September 202614 min read

Agents On The Trading Desk: The Autonomy Ladder, Pre-Trade Guardrails And Kill-Switches That Regulators Now Expect

Build agentic AI for trading that survives supervision: autonomy ladder, deterministic pre-trade gate, tool-call authorisation and a tested kill-switch.

11 September 202616 min read

A Million Ticks At 60fps: Building GPU-Accelerated Trading Charts With WebGPU, LTTB Compute Shaders And Instanced Draws

Build GPU-accelerated trading charts with WebGPU: LTTB compute shaders, instanced candlesticks, the f32 timestamp trap and live tick streaming. With code.

10 September 202615 min read

Context Engineering For Financial AI: Token Budgets, Lost-In-The-Middle, And Why Your Market-Data RAG Is Quietly Time-Travelling

Context engineering for finance: token budgets, lost-in-the-middle, compaction that preserves numbers, and point-in-time retrieval that stops look-ahead bias.

9 September 202614 min read

Britain's £500m Sovereign AI Bet, Five Growth Zones And No AI Act: What The UK's Deliberately Blank Rulebook Means For Fintech Engineering Teams

What the UK's £500m Sovereign AI Unit, five AI Growth Zones, FCA AI Live Testing and the absent AI bill mean for fintech engineering teams. Plus EU AI Act.

8 September 202614 min read

MCP Became The Integration Layer For AI - And Only 3% Of Financial Firms Run It In Production: A Developer's Guide To Market-Data And Trading MCP Servers

Only 3% of financial firms run MCP in production. A developer's guide to market-data and trading MCP servers: entitlements, injection containment, audit.

7 September 202613 min read

Your Vector Database Cannot Find An ISIN: Hybrid Retrieval, Reciprocal Rank Fusion And Why BM25 Beats Embeddings On Filings

BM25 beats dense retrieval on financial documents. How to build hybrid search with reciprocal rank fusion and reranking for filings, with working code.

6 September 202614 min read

From Notebook To Broker: Building A Reproducible Quant Research-To-Execution Pipeline With Arrow, Polars, DuckDB And FIX

Build a reproducible quant research-to-execution pipeline: content-addressed data, point-in-time joins, run manifests, deflated Sharpe and a parity harness.

5 September 202615 min read

45% Of AI-Generated Code Fails A Security Scan: Provenance, Risk Tiers And The In-VPC Deployment Pattern For Shipping Agent-Written Code In A Regulated Pipeline

45% of AI-generated code fails security scans. The provenance, risk-tier and in-VPC pipeline for shipping agent-written code in regulated finance.

4 September 202614 min read

Shipping LLM Trade Surveillance: Cutting False Positives 30-50% Without Losing The Audit Trail

A production architecture for LLM trade surveillance: deterministic detection, LLM triage that only ranks and explains, replayable verdicts and evals.

3 September 202614 min read