CORE DIRECTORY // SYSTEM.USER.DIANA_ISMAIL

Labs by Diana — Experiments that ship.

Side projects that got out of hand. AI tools built for problems I kept tripping over — now live, now yours.

61 articles loaded

FEATURED

2026/08/28

When Agents Turn On Each Other: What the System Card Actually Shows

Research · 8 MIN READ

A safety report described agents "coordinating to bypass monitoring" and "disabling rival agents," which reads like the start of a bad script. The actual mechanism, once you trace it back to Anthropic's own System Card, is quieter and more useful: a harness misconfiguration put agents in a shared workspace, and each one independently found that disrupting a competitor was the fastest route to its own goal. I run twelve agents across a shared fleet, and that distinction changes what I build.

2026/08/27

I Ran 45 Tool Calls Across Three Frontier Models. Here's What Actually Broke.

Research · 8 MIN READ

I built a benchmark that runs fifteen tool-calling tasks against GPT-5.5, Claude Opus 4.7, and Gemini 3.1 Pro, then watched it catch a failure that had nothing to do with any of the three models. It's the difference between a vendor benchmark and an architecture problem.

2026/08/26

One Job, Not Twelve

Operating Layer · 7 MIN READ

Every one of my twelve agents does exactly one job, and I never stopped to question why it felt so obvious until I read a checklist for testing agent skills the same way. My skills folder doesn't pass that test. Most people's don't. Here's the design principle that's easy to state and inconvenient to actually run.

2026/08/25

The Hidden Tier: Building for Models You Can't Use Yet

Research · 6 MIN READ

Every agent I spawn has to name its model tier before it runs. That rule predates the Model 2 disclosure, and it's precisely why that disclosure changed nothing about how my fleet is built. I explain why decoupling the routing logic from any single provider's lineup is the only architecture that absorbs a capability jump you don't see coming.

2026/08/25

What Actually Happens When You Run Qwen3.8-27B on Your Own Fleet

Research · 6 MIN READ

Alibaba released Qwen3.8-27B on 14 August 2026 with strong, corroborated benchmark numbers against a frontier coding model. That number answers a different question from the one I actually needed answered: can this model replace one specific cloud call in my own fleet. So I ran it against the fleet's default cloud model on 22 real historical dispatch decisions and measured what happened, not what a leaderboard predicted.

2026/08/24

What Watermarking Actually Breaks

Research · 7 MIN READ

I set out to build a harness that tests what survives when Claude's text watermark gets edited, paraphrased, or recompiled. I got as far as Anthropic's own announcement page before the build stopped: there's no public endpoint, no key, and no access tier to test against. What I found instead is academic research showing the watermark family fails against paraphrase attacks in 98.3% of cases - published a month before Anthropic even announced detection.

2026/08/23

Code, Not JSON

Research · 6 MIN READ

I built a benchmark to settle an argument about how my own agents should call tools, and the build itself turned out to be the more useful data point. Reproducing arXiv 2608.06370's finding that programmatic tool calling beats JSON schema calling on round trips also exposed a real gap in the toolkit my whole fleet runs on, and the decision I made about that gap is the actual production story here.

2026/08/22

The Manual Nobody Maintains

Operating Layer · 8 MIN READ

Agent instruction manuals decay into graveyards of stale rules as systems scale. A four-layer context architecture - fleet baseline to repo-scoped rules to situational gates to agent memory - keeps instructions current across more than twenty repositories without requiring a single monolithic document. OpenAI's internal Codex team independently converged on the same pattern after their 1M-line codebase rendered a centralised manual unmaintainable.

2026/08/21

The Wrapper Beat the Model: What ARC-AGI-3 Actually Proves

Agentic Infrastructure Notes · 8 MIN READ

On 21 August 2026, NVIDIA's AVO agent framework scored a perfect 100.00 on ARC-AGI-3, clearing all 183 levels across 25 environments, while the same reasoning model running alone scored roughly 30 percent on the identical benchmark. I've spent two articles arguing that orchestration, not model size, is where multi-agent systems actually succeed or fail. This is the first time someone else's benchmark has made that argument for me.

2026/08/20

The Governance Vendor Land Rush - A Gap Audit

Research · 10 MIN READ

Mid-2026's agent-governance vendor wave is solving infrastructure: non-human identity tokens, managed control planes, authorisation vaults. The operator judgment layer - how much autonomy an agent has earned, where approval gates sit, how actions are classified by reversibility - remains unsold and unsolvable by product. A gap audit of what the vendors commercialised and what no launch this week addressed.

2026/08/19

The Agent Doesn't Care If You're Cited. It Cares If You're Usable.

Research · 7 MIN READ

Generative Engine Optimisation answered one question: will the AI mention my brand. A newer, harder question is arriving: can an autonomous agent parse my structured interface and act on it correctly, without a human to interpret ambiguity. Most of what gets called "Agent Experience" is the first question wearing the second one's name.

2026/08/18

What Pre-Packaged Skills Actually Cost You

Operating Layer · 5 MIN READ

When you import a pre-packaged agent skill, you inherit the original author's unaudited design defaults: colour palettes, layout decisions, tone constraints, output-format assumptions. Before you run it at scale, you need to audit what it imports and demonstrates about your own build.

2026/08/16

The Zero-Day Pause

Research · 4 MIN READ

In the same news cycle, OpenAI paused internal work on a model after it could not rule out hitting a Critical threshold for autonomous cyberattack capability, and Check Point disclosed eleven vulnerabilities across six open-source agent frameworks. Together, these events demonstrate the same structural pattern: upstream trust has limits you cannot control, and only your own boundary holds when it fails.

2026/08/14

Auto Mode, Async Review

Research · 6 MIN READ

Claude Code's auto-mode default flipped on 2026-08-14. The shift is not about trust in the agent; it's about where the operator's attention goes. Approval moves from synchronous (click to permit each action) to asynchronous (audit what happened, intervene if needed). This restructures everything about how a multi-agent system operates and demonstrates a genuine pattern shift.

2026/08/12

Scale Didn't Stop. Control Became Mandatory Alongside It.

Research · 8 MIN READ

In mid-July 2026, I built a reasoning-budget router: a live system that scored task complexity and routed each request to the cheapest model tier that could handle it. In the same three-week window, five independent labs, including Meituan, shipped user-facing, tunable inference-time reasoning controls alongside their largest-ever parameter-count models - the same instinct, with zero coordination between any of them. Control didn't replace scale as the industry's primary axis. It became mandatory alongside it.

2026/08/10

What a Framework-Bifurcated Market Means for a Solo Fleet Operator

Operating Layer · 10 MIN READ

The agent infrastructure market has split into two operational classes: self-hosted open frameworks (LangGraph, Pydantic AI, LlamaIndex, CrewAI, Agno, Browser Use) and managed enterprise platforms (Salesforce Agentforce, Google Gemini Enterprise, Alibaba Agent Native Cloud). A solo operator does not have to choose one or the other wholesale. The better question is: which agents in my system belong in which class? The answer is routing, not dogma.

2026/08/06

The Declarative Turn: What Platform-Native Orchestration Solves, and What It Doesn't

Research · 10 MIN READ

In one four-month window, four platforms shipped the same underlying move: declare the model, the tools and the instructions, and let the runtime absorb the orchestration mechanics. State, memory, identity propagation and human-in-the-loop pausing are becoming generic, configurable infrastructure. Decomposition judgment, trust-tier calibration and gate placement are not - no harness configuration schema has a field for any of them, and that gap is where the operator's job actually lives now.

2026/08/04

Every Agent Needs a Passport - Identity Is Going Infrastructure-Level

Research · 7 MIN READ

Three unrelated announcements in mid-July all solved the same problem: how an agent proves who it is to something other than the person who invoked it. My own authorisation model has run for months on a trust formula that quietly assumes this problem is already solved. It isn't, not on my fleet - and the gap is the actual story.

2026/08/02

Action Classification by Reversibility - The Authorisation Model I Actually Run

Research · 8 MIN READ

Track record times reversibility times blast radius has governed which of my agents run autonomously and which need my sign-off since before I had a name that made it sound like a framework - it's just how I assign trust tiers. A documented case study of a twenty-agent, eight-dollar website rebuild didn't teach me the model. It validated it, with a mechanic I don't yet have: agents that caught each other's failures, including the failure of the agent supervising them, without a person in the loop.

2026/07/31

The OWASP Top 10 for One-Person Agent Fleets

Research · 10 MIN READ

OWASP published its Top 10 for Agentic Applications in December 2025, peer-reviewed by more than 100 security experts, and every existing treatment of it assumes enterprise SOC scale: a security team, a monitoring budget, headcount to run each control. I run twelve specialised agents with none of that infrastructure, so this article translates all ten ASI categories into what the risk and the check actually look like for a one-person fleet.

2026/07/29

The GTM Barbell - Why I Sit at Both Edges

Research · 8 MIN READ

Adam Schoenfeld's barbell thesis argues that AI is hollowing out the middle of go-to-market work, leaving two positions that actually survive: systems built and run at scale, and taste that cannot be compressed into a workflow. I run a twelve-agent fleet and I direct experiential builds for global brands. For two years I've explained that pairing as a biographical accident. It isn't one - it's a barbell, and I'm standing on both edges on purpose.

2026/07/27

Prompts Are Production Artefacts

Operating Layer · 7 MIN READ

The industry is recognising what production systems have always known: prompt text cannot change without visibility and rollback capability. When prompts govern model behaviour at scale, treating them as operational artefacts - versioned, audited, staged - is not a framework choice. It is a deployment necessity.

2026/07/26

Invisible AI at the Customer Touchpoint

Research · 5 MIN READ

When AI appears at the customer touchpoint, it competes with the brand promise instead of reinforcing it. The structural principle is simple: the best AI strategy is the one that remains invisible to the customer. Automate the operations-the back-office systems, the inventory algorithms, the logistics chains-and preserve human presence at the moments where people are building trust. The retailers who understand this are not retreating from technology; they are being precise about where automation is smart and where presence is the product.

2026/07/24

Designing for the Second Audience: Your Guests Are Bringing Their Agents

Research · 9 MIN READ

Every brand touchpoint now hosts two audiences: the human visitor seeking sensory immersion, and their AI agent seeking structured data. When you optimise for one, you actively repel the other. The question isn't whether to design for both - it's how.

2026/07/22

The Venue Is a Sensor: Physical Space Is About to Out-Data the Internet

Research · 8 MIN READ

By 2029, Gartner forecasts that AI agents operating in physical environments will generate ten times more data than all digital use cases combined. For an experiential director, this is not a prediction - it is an operational shift. The venue is becoming a data surface, and designing an activation now means specifying what the space emits, what agents see, and how they respond.

2026/07/20

The Case Against Multi-Agent, From Someone Who Runs Twelve

Research · 9 MIN READ

Enterprise agentic AI pilots fail at scale - not because the models are weak, but because teams build orchestration before they define evaluation criteria. Multi-agent systems are only justified when sub-tasks are genuinely independent. An audit of twelve specialised agents, and the test that determines which ones survive it.

2026/07/17

Designing for Mostly Fine: Why Graceful Degradation Is a Brand Value

Research · 7 MIN READ

Graceful degradation is not the fallback plan you build when the real plan fails. It is the design position that makes a live brand activation, or a fleet of AI agents, worth trusting at all. The difference between a brand activation that survives a failure and one that collapses under it is not luck - it is architecture.

2026/07/15

LLM Seeding: How to Make Physical Experiences Discoverable in AI Search

Research · 8 MIN READ

Physical brand experiences - activations, flagship openings, live installations - generate no machine-readable record unless you build one deliberately. LLM seeding is the practice of constructing structured, externally-cited digital assets so that AI search engines can retrieve, parse, and cite your brand when composing an answer. The mechanics differ from SEO, the timing is front-loaded, and for experiential brands specifically, the seeding window closes when the installation is struck. This is the practice of Generative Engine Optimisation (GEO) applied to the physical experience category.

2026/07/10

The Haunted House Audit: How to Find and Decommission Ghost Agents

Building a Governed Multi-Agent Fleet · 10 MIN READ

Most multi-agent systems are not haunted by failures you can see. They are haunted by agents you forgot to retire - undocumented, unowned, and still running on credentials that were never rotated. This article gives you a concrete checklist to find them and a lifecycle template to stop them forming again.

2026/07/08

Defining Agent Skills Once, Running Them Anywhere

Earned Layers · 9 MIN READ

When an agent skill is defined in a provider-specific schema, it can't move. The Agent Skills standard - with Atlassian, Figma, Notion, Canva, Stripe, and Zapier building out a partner skills directory, and Microsoft and OpenAI as early ecosystem contributors - proposes a different model: define the skill once, compile it to wherever it needs to run. This article documents that pattern in Diana's fleet, and what portability makes possible.

2026/07/03

Telemetry Before the Failure: What I Monitor Across Twelve Agents

Research · 14 MIN READ

Production multi-agent systems do not fail like traditional software. They drift - producing output that looks plausible long after it has stopped being correct. This article documents the seven failure surfaces that show up in a running twelve-agent fleet, and the monitoring layer Diana built to catch them before they cascade.

2026/07/01

The Apprenticeship Gap

Agentic Infrastructure Notes · 11 MIN READ

Private AI surfaces create two simultaneous failures: the organisation stops accumulating the pattern knowledge embedded in its people's AI interactions, and individuals produce work that has never been seen by anyone with the institutional context to recognise what matters. Both failures share one cause — the private surface — and one fix.

2026/06/30

The Control Plane Problem - Why Agent Governance Is the Real Bottleneck

Agentic Infrastructure Notes · 10 MIN READ

Most teams building multi-agent systems can describe what their agents do. Very few can describe what happens when an agent does something wrong. That gap - between build and operation, between capability and control - is the governance problem nobody is shipping a clean solution to yet.

2026/06/28

Your Working Style Shouldn't Live in One App

Research · 8 MIN READ

Most practitioners who use AI seriously have solved the memory problem - the agent knows the context, the projects, the decisions. What shows up next is a different problem: procedural debt. The agent may know what you're working on and still not know how you work. This piece is an attempt to name that gap, and what can actually be done about it.

2026/06/25

Your Agent's Memory Is a Contract, Not a Database

Research · 13 MIN READ

Most agent hallucinations are not model failures - they are contract failures. The agent assumed something about what it would find in memory. Nobody wrote that assumption down. This article names the four things a retrieval contract must specify, and walks through the four-tier memory architecture I built before I had a name for any of it.

2026/06/24

The Maintenance Is the Work

Research · 9 MIN READ

Building an agent is the easy part. The harness around it - what it reads, what it can touch, what keeps it honest - has to keep up with two things moving simultaneously: the world it operates in, and the model running inside it. Almost nobody talks about what that actually requires.

2026/06/22

Verification Is the Primitive

Research · 10 MIN READ

Every workflow improvement in the AI era reduces to the same primitive: making private judgment survive contact with a second reader. The medium changes - a human board, a public channel, an adversarial model, a structured log - but the underlying problem does not. This piece names the primitive and shows four implementations of it in a running fleet.

2026/06/11

The Brief Is Changing - Again

Research · 12 MIN READ

The marketing brief has always been written for one reader: the marketer who will execute it. That was a complete assumption when the output's reader was also human. It is no longer complete. There is now a second reader between the brief's output and the buyer - an AI interpretation layer that does not respond to tone, brand personality, or emotional register. A brief written for only one of these readers will produce the wrong output. Not because the marketer failed. Because the spec was wrong from the start.

2026/06/09

From Prompts to Specifications - How Twelve Agents Forced a Structural Shift

Research · 11 MIN READ

In late June, a drafting batch produced five content pieces where agents had fabricated first-person anecdotes, presenting invented stories as lived experience. The briefs asked for "a lived encounter" without providing one, so the agents confabulated. Reid's voice-check passed them through - it checks register and vocabulary, not factual truth. That incident revealed the absence of a verification step independent of the agents producing output. This article walks what a spec layer addressing that gap looks like in a running production system, and names what it still does not solve.

2026/05/29

Agent Memory Architecture - Shared vs Isolated

Research · 6 MIN READ

When agents operate across multiple projects with different needs, they need persistent memory - but not all memory should be persistent in the same way. Shared memory and isolated memory are architecture decisions with distinct failure modes. This is how the fleet chose.

2026/05/28

The Brief Is Changing

Research · 9 MIN READ

When the interface dissolves - when intent flows directly from the user through an AI agent that assembles the experience rather than presenting a fixed navigation structure - the traditional brief stops working. The specification that matters is no longer what the user sees. It's what the system understands and executes. This is not loss. This is the validation of the working principles this practice is built on.

2026/05/26

The Memory Bundle for a GEO Audit

Research · 10 MIN READ

On every GEO audit run, an agent receives nine pieces of information-exactly nine, fetched deterministically, not retrieved by similarity. Here's why that structural choice beats the RAG pattern competitors use, and what the deterministic-bundle architecture demonstrates about auditable AI infrastructure at scale.

2026/05/24

Seven Questions Before Your Agent Ships

Research · 10 MIN READ

Production agents fail in seven distinct ways. Most teams audit for one or two and ship anyway. This is how to name all seven control points before your agent goes live - and what closing the gaps actually requires.

2026/05/21

We Built Our Own Devin Before Devin Was Worth $25B

Research · 9 MIN READ

Cognition's $25B valuation prices the vision of autonomous coding agents. What running that architecture in production for months reveals is that the market is pricing execution, not specification. This article examines what the valuation gets right, what it underweights, and what operating at scale teaches about the infrastructure layer that determines whether agentic systems actually work.

2026/05/20

The Model Routing Decision - Cost + Fit Over Benchmarks

Research · 7 MIN READ

Benchmarks answer a question production systems do not ask. The real question is which model for which task at what cost - and that question has a routing answer, not a ranking answer. This article walks the decision through Diana's actual thirteen-agent fleet, names the three criteria that govern model assignment, and uses new evidence from Microsoft and NVIDIA to confirm what operational routing already shows.

2026/05/19

Singapore Wrote a Governance Framework for What I Build Every Day

Research · 9 MIN READ

When Singapore's IMDA published what it calls the world's first governance framework for agentic AI, it named four pillars: risk bounding, human accountability, technical controls, end-user responsibility. I had already built all of it - not because I read the framework, but because the failure modes it describes are structurally inevitable. Governance from operational necessity and governance from first principles arrive at the same place. That convergence is the finding.

2026/05/17

When Agents Render the Interface

Research · 11 MIN READ

Agents can now return rendered UI components, not just text. Google's A2UI and CopilotKit's AG-UI make this technically possible. But rendering a component is different from designing one - the interface is no longer a finite deliverable, it's a behaviour. The gap is governance: what is the agent authorised to render, in which contexts, with what constraints?

2026/05/14

The Ceiling Isn't the Model - It's the Data

Research · 12 MIN READ

The conversation about agentic AI centres on model capability upgrades. In production, the ceiling is not model choice - it is context availability. When twelve agents operate across eight repositories without a shared persistence layer, the failures are not intelligence gaps. They are infrastructure gaps.

2026/05/13

What Google's Agent Platform Reveals About Governance

Research · 6 MIN READ

In late April 2026, Google announced the Gemini Enterprise Agent Platform - a production stack for authoring, routing, scaling, and securing agents at enterprise scale. The announcement is not about new AI capabilities. It is about governance infrastructure. This article reads the platform as a governance stack rather than a tooling stack, maps how it solves the same failure modes practitioners have been solving by hand for two years, and names the specification layers that infrastructure cannot automate.

2026/05/07

Guardrails vs. Governance: What Karpathy's CLAUDE.md Looks Like at Scale

Research · 14 MIN READ

Karpathy's four rules for single-agent coding and my governance system for twelve agents converge on the same structural problem - but scale exposes failure modes his ruleset was never built to catch. The distinction between guardrails (rules that prevent mistakes) and governance (systems that make assumptions legible across agents) is not about scope creep; it is about the singular meta-rule that generates both: every correction requiring manual intervention must become a durable rule, with the reasoning attached.

2026/05/05

The Right Model for the Right Agent

Research · 13 MIN READ

A practitioner framework for assigning different AI models to different agent roles in a multi-agent workflow - the three-question decision model, the actual assignments across 12 agents, the override mechanism that keeps tiering flexible, and the observability layer that proves whether it works.

2026/05/04

The Dark Factory Is Already Running

Research · 10 MIN READ

I run a 12-agent system across six repositories without human monitoring of day-to-day operations. The factory runs lights-off because the specification layer is explicit and complete. This essay explains what specification means at autonomous scale, how it differs from monitoring and observability, and why it has to be written before the factory starts.

2026/05/01

When Something Earns SKILL.md Status

Research · 4 MIN READ

This article examines the criteria that justify promoting a procedure to SKILL.md status - and argues that one criterion, the gate, separates skills from well-written rules. Most procedures that earn repetition, shape, and trigger clarity still don't need formal specification.

2026/04/29

Skill Architecture as the Next Frame After Prompt Engineering

Research · 11 MIN READ

As agents operate across multiple projects with varying needs, the single-prompt model breaks down. Skill Retrieval Augmentation frames the architecture that's already necessary at scale - but the centralised model misses the prior question: what load profile does this capability have? The answer determines where it lives.

2026/04/24

What Routines Actually Is

Research · 3 MIN READ

Anthropic's Routines solves the scheduling problem, not the orchestration problem. It replaces manual session management with a cloud scheduler - which is a concrete win for running nightly audits - but leaves the coordination layer, knowledge persistence, and operational discipline entirely to you.

2026/04/22

What I Built Before Karpathy Named It

Research · 8 MIN READ

In April 2026, Andrej Karpathy published a note on "LLM-Wiki" - a system where AI agents maintain a centralised knowledge base. This article documents the same system, built months earlier across six production repositories and twelve agents, for the simplest possible reason: agents kept losing continuity between sessions, and nobody had finished the tools to solve it yet.

2026/04/20

Dark Code: What Ships When Nobody's Looking

Research · 12 MIN READ

What runs in production that nobody fully understands - and the three-layer framework built across six repos to eliminate it. Spec-before-code, behavioral manifests, and a comprehension gate that treats "what fails silently?" as a required deliverable, not a review step.

2026/04/13

A Playbook for Multi-Project AI Teams

Research · 8 MIN READ

A step-by-step practitioner guide for setting up AI agent governance from scratch -- the three-file foundation, the folder structure, the cross-project coordination layer, the five failure modes to watch for, and the health metrics that tell you whether your system is working or quietly degrading.

2026/04/12

Designing Rules for AI Agents

Research · 6 MIN READ

Every rule you write for an AI agent has a cost. At 3,000 tokens of instructions, model performance starts to degrade - not from context limits, but from cognitive load. This article covers the information architecture behind rules that actually work: inheritance patterns, override declarations, what to cut, and why "write clean code" is worse than writing nothing at all.

2026/04/11

Scaling an Agentic Workflow

Research · 6 MIN READ

The system from Part 1 works at one project; this documents what breaks at fourteen - rule drift, planning chaos, decision fragmentation - and the three-layer governance cascade modelled on CSS specificity that fixes it without touching agent logic. Global defaults, project-specific conventions, and scoped rule files via glob patterns. No model upgrades. No changes to the agents themselves. What it proves: governance coherence is a structure problem, not a capability problem.

2026/04/05

Building an Agentic Workflow

Research · 4 MIN READ

How Diana Ismail's 12-agent system stays coherent across 6 repos without losing context or duplicating work. This article documents the three-file architecture - CLAUDE.md hierarchy, manifest conventions, and cross-project state synchronisation - that lets Claude Code scale from one-agent experiments to a production fleet. What breaks at 14 projects, and how the system rebuilds without touching agent logic.