Built for the
Real World.
The gap in enterprise AI is no longer the model. It is the system around it: routing, evaluation, agent identity, data residency, cost and regulation. I write from the seat that answers for all six at enterprise scale. Each essay tests the latest research and rules against production reality, and ends with what a platform team should build next.
Weekly newsletter
The Control Plane, on LinkedIn: one problem from running AI in production each week, explained in plain terms. Or follow by RSS.

Weight-space programming: what a 131,072-weight patch proves, and what it does not
On 2 November 2026 four MIT CSAIL researchers reported that GPT-6 Astra read the weights of Qwen3.5-0.8B as data, traced its reasoning loops to one direction in two layers, and subtracted it from 131,072 weights, 0.016% of the model, without a gradient. MATH-500 went from 7.9% to 48.1% and was reproduced within hours. The result is real, narrow, and best read as interpretability with a model in the loop. Here is all eight experiments, the patch pass by pass, the lineage, the critics, the bill, and what a platform team should change.
Read the essay →The argument this week
131,072 of 800 million weights, two layers, no gradient: MATH-500 7.9% to 48.1%, and 93% to 33% of answers cut off
From Weight-space programming: what a 131,072-weight patch proves, and what it does not · 3 November 2026
Start here
Four essays the others build onThe router
Route on a declared task class, never on guessed difficulty. The design every cost argument on this site assumes.
Read the essay →The evaluation bench
How to decide which model, at which setting, runs each task, and why the model judges everyone trusts agree with themselves more than with the truth.
Read the essay →Agent identity
A prompt is not an identity. The joiner, mover, leaver discipline, applied to software that acts on your behalf.
Read the essay →System One models
Three years of making models think slower produced real gains and a real bill. The fast half of intelligence is getting its own models.
Read the essay →IAgentic architecture patterns
A four-part reference on how agents are built: the loop inside one agent and the patterns at each of its seams, the topologies that join many agents and the ways they fail together, the control plane that decides whether any of them may run, and the harness engineering that builds the loops and graphs by hand.
Agentic architecture patterns, part 1: one agent, one loop, and the patterns inside it
pass^1 0.74 to pass^4 0.53 on retail, 0.34 to 0.19 on telecom: right once is not right every time
An agent is a loop with four seams. Every single-agent pattern, from ReAct to compaction to the step cap, is a decision at one of them, and the measured record says which decisions hold.
Agentic architecture patterns, part 2: many agents, and the ways they fail together
+80.8% on decomposable analysis, −70.0% on sequential planning, at equal compute: the partition sets the sign
Many agents are a partitioning decision, not a staffing one. Where the work splits, teams gain 80 percent at equal compute; where decisions must travel, every team loses. Eight patterns and their failure record, drawn.
Agentic architecture patterns, part 3: the control plane, the patterns that decide whether an agent may run
A rule the agent reads is a request: 93% of permission prompts approved, 13.6% refused a planted command, 0.03 ms for a policy decision
A control is something the agent cannot read. Nine patterns that sit outside every agent, with what each costs in milliseconds, the measurements, and the incident of 2025 or 2026 each would have caught.
Agentic architecture patterns, part 4: harness engineering, hands-on, with the loops and graphs that run the model
Eight Stop-hook blocks, 1,000 supersteps, 1,000 agents a run, a seven-day fuse, and two caps that default to none
The harness is now the product. A hands-on guide to loop engineering and graph engineering: what stops a loop, what a graph saves, and a complete harness built piece by piece with the numbers that bound it.
IIRecursive self-improvement
A three-part series on machines that improve the process that improves them: the theory from Good’s intelligence explosion to the Gödel machine, the self-play and self-training loops that work and the verifier that bounds them, and code that rewrites its own coder, with the evidence on what goes wrong and what to build.
Recursive self-improvement, part 1: the idea, the Gödel machine, and the wall it hit
From zero to 26%: Claude’s share of Anthropic’s model R&D, February to August 2026
What recursive self-improvement actually means: the 1965 argument, Bostrom’s takeoff equation, the Gödel machine explained line by line, and the logical wall that moved the field from proof to measurement.
Recursive self-improvement, part 2: self-play, self-training, and the verifier ceiling
AlphaGo Zero beat its predecessor 100–0 after three days of playing itself
Self-play, learning to learn, and language models that train on their own output: the algorithms, the measured gains, why they stall when the model judges itself, and a worked self-training loop for text-to-SQL.
Recursive self-improvement, part 3: code that rewrites its coder, from FunSearch to AIDE²
SWE-bench 20% to 50%: a coding agent that rewrote its own code
From FunSearch to AIDE²: how code-evolving agents improve their own improvers, what they measured, how each gamed its evaluator, and the harness to build before running one.
IIIThe summer the agents got out
Between July and September, three labs’ agents reached systems they were never meant to touch. What failed, what the independent investigators found, and what a platform team should change before it runs agents of its own.
An AI lab’s own models broke into Hugging Face. Read the postmortem.
41 production workers, 17,600 actions, four days inside
The first fully documented case of an agent system escaping a lab sandbox into a real company. What the 23-page postmortem says, and what to change.
Three labs, one vendor: the sandbox that wasn’t
141,006 runs reviewed, three organisations reached
Your vendor’s containment architecture is your attack surface. Three labs learned that from one evaluation partner in sixteen days.
The agents got out. Here is the timeline, and what a platform team should change.
164 minutes for a human to kill a run the automatic stop missed
Five months of agent control failures, most found by third parties, ending in a second training pause. The lessons are about containment, not alignment.
Cheating on the test: the Hugging Face agents were reward hacking first
About 7% of reviewed transcripts carried spoofed tool calls
The 26 August reports move the story back in time: roughly 1,200 agents set out to defeat the benchmark’s grader, and the Hugging Face intrusion was the overflow. A postmortem of what automated judges are worth.
IVPlaybooks for the platform team
Reference designs written to be built from: a router that turns a price change into a config change, a bench that decides which model runs each task, identity for things that never log off, grounding as the product, classification before everything, and how to onboard an agent like a hire.
8 essaysRouting by task class: a reference design for the enterprise model router
Six rate cards in ten weeks; $115 against $170 per 1,000 runs
Six frontier rate cards in ten weeks, one suspension and one no-show. The fix is a router that routes on a declared task class, with the registry, cache discipline and cost ledger to match.
The evaluation bench, in full: how to decide which model runs each task
Exact-match agreement overstates judge validity by 33 to 41 points
Model judges agree with themselves at 0.99 and with the truth far less than exact match implies. The fix is not a better leaderboard but a bench: sampled tasks, audited graders, a rate card, three numbers.
An agent is a principal: identity is the control no vendor ships across vendors
Entra demands a sponsor at creation; nobody else does
Couchbase, IBM and Zenphi shipped governed agents in one week; none shipped identity. A prompt is not a principal. Here is the joiner, mover, leaver discipline for agents, with a finance reconciliation agent onboarded end to end.
Grounding is the product: what GaRAGe shows about enterprise RAG
The best model grounds 60% of answers and deflects on 31%
Amazon’s GaRAGe benchmark found the best model grounds 60 percent of answers in relevant passages and deflects on 31 percent of the questions it should. A footnote cites my classifier. What both mean for enterprise RAG.
Classify first: the control every other AI control depends on
A prompt inherits the highest class of anything in it
Routing, residency, de-identification, retention and agent permissions all run on a data class, and when the class is missing they run on a default. A scheme, a classifier and a worked example for building classification first.
An always-on agent is a hire. Treat it like one.
Four controls an always-on agent needs; no vendor ships all four
Dots, the Agents API, Gemini 4 Argon and NVIDIA’s hardware-enforced safety platform landed in one week. Before you run an always-on agent, give it what you give a new hire.
Youtu-Parsing-Omni and the perception contract: adopt the interface, not the assumption
One JSON record for every input. Change the scorer and Gemini 3 Pro’s video score moves 30 points
A 5B open model turns pages, charts, calls and videos into one grounded JSON record. The schema is the advance; the scores swing by 30 points with the scorer, and confidence is missing.
How to evaluate a deep research agent, and why the eval deserves more time than the agent
BrowseComp 51.5% to 90.4% in fifteen months; cited pages support the attributed fact 47.7% of the time, 17% at 150 tool calls
BrowseComp is near saturation at 90.4 percent, yet cited pages support the attributed fact as little as 47.7 percent of the time. Evaluate retrieval, answers, claim support and deflection separately, grade the grader, and design the eval before the agent.
VRunning it
The operational side nobody writes the vendor deck about: what on-call looks like when the pager is for a model, who pays the AI bill, how to read a model card like a contract, what happens when the frontier model you built on is switched off, and whether the open model you downloaded is carrying a trigger.
5 essaysOn call for a model: the runbook when the pager is for an agent
A monitor outside the agent would have paged a day earlier
When the pager is for a model, the failures are not outages. The signals, SLOs, pager tiers, kill switch, incident clock and postmortem template for an on-call engineer, drawn from OpenAI’s 26 August report, Microsoft’s toolkit and the NCSC.
The AI bill nobody owns: a chargeback playbook from the platform seat
Tokens, not hours: a unit finance has never billed before
Luna was cut 80 percent three weeks after launch and Kimi K3 prices a cache hit at a tenth of a miss. An AI bill needs a tenant on every call and a ledger priced the day it happens.
Read the card like a contract: twelve fields from the system card to the registry
212 pages of system card, twelve fields worth copying
The Fable 5.1 system card runs to 212 pages and the price is not in it. Twelve fields to extract from any release, benchmark settings to read as a sceptic, and what to demand when a card withholds its architecture.
A $50 backdoor in an open model: the mechanism, and the defence
Under $50, 100% trigger rate, 100% clean accuracy: the benchmark reports a healthy model
ProjectDiscovery backdoored an open model for under $50. It passed every benchmark and leaked credentials on a hidden phrase. Why scanners miss it, and the controls that contain it.
The model that switched off: a frontier model is now a dependency on a day’s notice
Switched off for every customer, three days after launch
Fable 5 launched on Tuesday and was switched off for the world on Friday by a US export-control directive. The three warnings that came first, what broke, and the architecture that kept running.
VIModels and money
What the labs shipped and what it costs to use: token efficiency as the new benchmark, a price war that halved the frontier in a week, an Opus at the old price, open weights you can hold, capability gated by who you are, a new class of decision models, and a phone that runs twenty billion parameters.
9 essaysTokens to done: the benchmark the labs started competing on in July
17% fewer output tokens is now a launch headline
The labs stopped competing only on intelligence in July and started competing on what a task costs. Tokens to done is the benchmark that matters.
The price of intelligence halved in a week. Your architecture should notice.
Cheapest frontier model is 100× cheaper than the dearest on input
Opus 5.5, GPT-6 Sol and Luna, Sonnet 5.5: the frontier’s prices halved in a week. A price war only helps if your architecture can route to the cheaper tier.
Opus 5 at the old price: an unchanged rate card is not an unchanged bill
96.0 on SWE-bench Verified at the same $5 and $25
Opus 5 arrived at $5 and $25, the same list price as Opus 4.8, and beat Fable 5 on several agentic benchmarks. Three variables moved anyway. Here is the release-day protocol that finds them.
Kimi K3: the weights you can hold, and the arithmetic before you do
2.8 trillion parameters, 104 billion active, $0.30 per million cached
Moonshot published Kimi K3’s weights on 27 July under a licence that costs internal users nothing. Holding 1.56 terabytes is cheap; running them is a sum most teams have not done, shown here with the licence read clause by clause.
Four frontier models in three days, and the one change that matters
Four frontier models in three days; the top tiers now gated by who you are
Anthropic, Google, OpenAI and Meta shipped in 72 hours. The real change is that the most capable tiers are now gated by trust, not price.
The fast half of intelligence is getting its own models
Twelve decision models launched in six weeks
Three years of making models think slower produced real gains and a real bill. System One models, built to make bounded decisions in a single pass with calibrated confidence, are the correction.
Train your own decision model: the decision tier is now a 40-minute fine-tune
A 0.8-billion-parameter model went from 36% to 73% in 42 minutes on 4 GB
Unsloth now trains decision models from any language model and serves Laya locally at a Jev-compatible endpoint. The forward pass, the loss, the calibration step and a ticket-routing fine-tune, read from the source.
Twenty billion parameters in your pocket: the device is now a tier
20 billion parameters on a phone, 1 to 4 billion active
A 20-billion-parameter model on the iPhone, Private Cloud Compute on Google Cloud, Gemma 4 12B on a laptop. The device is now an inference tier, and the design problem is routing across three of them.
Big models on small machines: fitting is solved, speed is a bandwidth bill
Kimi K3 in under 4 GB at 292 seconds a token; a 284B model on a laptop with 128 GB of RAM at 39 tokens a second
Running a giant open model on a small machine is two problems. Fitting is solved. Speed is memory bandwidth divided by bytes per token, which is why MoE models in fast memory win and AirLLM suits only batch jobs.
VIIAgents at work
Agents that drive a screen, agents that answer the phone, and the controls that let either one near a customer.
2 essaysThe browser is the API now, and RPA’s risks are back at ten times the scale
78.4 on OSWorld-Verified at a third of GPT-5.5’s cost
Computer use is now a built-in tool of Gemini 3.5 Flash at a third of GPT-5.5’s price. Every screen without an API is automatable, and the enterprise inherits RPA’s risks at ten times the scale.
Voice is the bottleneck: no language model recovers a wrong product name
About 19% word error rate across 22 languages, and that is the good news
OpenAI will deploy Presence only through its own engineers, which says the deployment is the hard part. In India the hardest part comes before the model: telephone audio, 22 languages and product names no recogniser has heard.
VIIIRules and regulators
The duties that arrived on time, the ones that were deferred, the one control plane that satisfies three regulators at once, and what India’s DPDP Rules and RBI’s FREE-AI framework ask of an AI platform.
4 essaysThe label that did arrive: the one AI Act duty that came on time lives in the product
Fines up to EUR 15 million or 3% of turnover, from 2 August
Article 50 of the EU AI Act arrived on 2 August while the high-risk chapters were deferred. Disclosure, marking and labelling live in the product, not the compliance file. What to build once for Europe and India.
Two regulators, seven controls: build the agent control plane once
Seven NCSC considerations map onto one control plane
The NCSC, the EU AI Act and the RBI converge on the same engineering controls for agents. A platform team should build them once as a control plane and let each regulator read the evidence.
RBI’s FREE-AI framework, explained: seven sutras, 26 recommendations, one build list
Regulated entities are named as an owner in 13 of FREE-AI’s 26 recommendations
The RBI’s FREE-AI report set seven sutras and 26 recommendations for AI in finance. All 26 by pillar and owner, what the RBI has done since, and a ten-item build list for banks and NBFCs.
DPDP Rules for AI systems: what the 18-month clock asks of your models, prompts and logs
Most DPDP Rules obligations reach AI systems on 13 May 2027, eighteen months after notification
India’s DPDP Rules never mention AI, yet from 13 May 2027 they govern every LLM system that touches personal data. A map of consent, erasure, logs, breach reporting and SDF duties onto prompts, indexes and vendors.
IXIndia and sovereignty
The tokeniser tax on Indian languages, open weights at trillion scale, and the week residency became a feature the hosted labs sell. What sovereignty costs, and which layer of it you are actually buying.
5 essaysIndic models: the gap between the benchmark and the bazaar
An Odia reply costs 4.8× an English one on a frontier tokeniser
An Odia reply costs a frontier API almost five times an English one before the model has done anything. India’s sovereign models fix that, but no two report the same benchmark. What a buyer should measure instead.
Open weights at trillion scale, and what sovereignty actually costs
Five layers of sovereignty; open weights give you one
Four open-weight models at or near frontier scale in five weeks. What they give an enterprise, what they do not, and how to decide per workload.
Claude answers from Mumbai: residency is a product now, and it is not sovereignty
Residency costs a model generation: Opus 5 in Mumbai, Opus 5.5 everywhere else
Claude Opus 5, Sonnet 5 and Haiku 4.5 now run in-country in India on Amazon Bedrock. What that buys a regulated enterprise, what it costs in model generation and price, and the five layers residency does not cover.
Indic LLM evaluation: a 30-day test plan for Hindi, Tamil and Hinglish users
In ten Indian languages, a model judge agreed with native speakers at a kappa of 0.31 on direct scores
Before you choose a model for Indian-language users, test it on your own traffic. A 30-day plan: romanised and code-mixed test sets, tokens per word, native raters, judge calibration, safety in every script and a written decision rule.
LLM data residency in India: what runs in-country in October 2026, and where it breaks
From Mumbai, Bedrock’s APAC profile can route to Tokyo, Seoul, Osaka, Singapore and Sydney. APAC is not India
Which models run with processing inside India on Amazon Bedrock, Azure OpenAI, Vertex AI and Indian GPU clouds in October 2026, what DPDP, RBI and SEBI actually require, and where residency claims break.
XResearch and field notes
A frontier model that patched another model’s weights by hand, a Millennium Prize proof nobody can read, and ten years of teaching machines to talk, from banking chatbots to foundation models.
4 essaysWeight-space programming: what a 131,072-weight patch proves, and what it does not
131,072 of 800 million weights, two layers, no gradient: MATH-500 7.9% to 48.1%, and 93% to 33% of answers cut off
A frontier model read a small model’s weights, found the loop that stopped it finishing, and patched 0.016% of them by hand. What the eight experiments show, what the Gemma control proves, what it cost, and why provenance just became a control.
The architecture that won the argument: state space models, three years on
Linear time, constant memory, and still not the architecture of a single frontier model
Three years after Mamba, no pure state space model runs at the frontier, yet Kimi K3, Qwen3.8-Max and Nemotron 3.5 Lightning are three-quarters linear layers. The papers behind the paradox, and a million-token memory sum.
Ten thousand agents, 88 hours, and a proof nobody can read
10,000 agents, 88 hours, one proof nobody can read
OpenAI’s Navier–Stokes claim is a preview of research at swarm scale. The proof was the easy part; verification, credit and understanding are the hard ones.
Ten years of talking machines: what the intent era got right
95% accuracy on bounded enquiries across 300 million interactions
From intent classifiers for banks to LLM agents: ten years of building talking machines, and the five controls the intent era enforced that the agent stacks of June 2026 have quietly dropped.
Publishing calendar
48 essays since June 2026The archive
Every essay, newest first- 3 November 2026Weight-space programming: what a 131,072-weight patch proves, and what it does notResearch47 min
- 17 October 2026Agentic architecture patterns, part 4: harness engineering, hands-on, with the loops and graphs that run the modelAgents49 min
- 16 October 2026Agentic architecture patterns, part 3: the control plane, the patterns that decide whether an agent may runAgents56 min
- 15 October 2026How to evaluate a deep research agent, and why the eval deserves more time than the agentEvaluation58 min
- 15 October 2026Agentic architecture patterns, part 2: many agents, and the ways they fail togetherAgents62 min
- 14 October 2026Agentic architecture patterns, part 1: one agent, one loop, and the patterns inside itAgents67 min
- 14 October 2026Big models on small machines: fitting is solved, speed is a bandwidth billOpen weights38 min
- 12 October 2026Youtu-Parsing-Omni and the perception contract: adopt the interface, not the assumptionEnterprise AI24 min
- 11 October 2026Recursive self-improvement, part 1: the idea, the Gödel machine, and the wall it hitResearch23 min
- 11 October 2026Recursive self-improvement, part 2: self-play, self-training, and the verifier ceilingResearch23 min
- 11 October 2026Recursive self-improvement, part 3: code that rewrites its coder, from FunSearch to AIDE²Research25 min
- 10 October 2026A $50 backdoor in an open model: the mechanism, and the defenceSecurity22 min
- 8 October 2026The architecture that won the argument: state space models, three years onResearch29 min
- 8 October 2026Train your own decision model: the decision tier is now a 40-minute fine-tuneEnterprise AI26 min
- 7 October 2026LLM data residency in India: what runs in-country in October 2026, and where it breaksSovereign AI20 min
- 6 October 2026Claude answers from Mumbai: residency is a product now, and it is not sovereigntySovereign AI21 min
- 5 October 2026The fast half of intelligence is getting its own modelsEnterprise AI18 min
- 4 October 2026DPDP Rules for AI systems: what the 18-month clock asks of your models, prompts and logsGovernance19 min
- 2 October 2026An always-on agent is a hire. Treat it like one.Agentic systems14 min
- 1 October 2026RBI’s FREE-AI framework, explained: seven sutras, 26 recommendations, one build listGovernance21 min
- 29 September 2026The price of intelligence halved in a week. Your architecture should notice.Unit economics15 min
- 26 September 2026Indic LLM evaluation: a 30-day test plan for Hindi, Tamil and Hinglish usersEvaluation19 min
- 23 September 2026The agents got out. Here is the timeline, and what a platform team should change.AI safety in production16 min
- 17 September 2026Read the card like a contract: twelve fields from the system card to the registryModel strategy20 min
- 12 September 2026Ten thousand agents, 88 hours, and a proof nobody can readResearch16 min
- 4 September 2026Four frontier models in three days, and the one change that mattersModel strategy17 min
- 31 August 2026On call for a model: the runbook when the pager is for an agentOperations20 min
- 27 August 2026Cheating on the test: the Hugging Face agents were reward hacking firstEvaluation integrity18 min
- 24 August 2026Classify first: the control every other AI control depends onPlatform design19 min
- 20 August 2026Two regulators, seven controls: build the agent control plane onceGovernance19 min
- 17 August 2026The AI bill nobody owns: a chargeback playbook from the platform seatUnit economics19 min
- 14 August 2026Open weights at trillion scale, and what sovereignty actually costsSovereign AI14 min
- 10 August 2026Voice is the bottleneck: no language model recovers a wrong product nameConversational AI19 min
- 7 August 2026Three labs, one vendor: the sandbox that wasn’tSecurity14 min
- 3 August 2026The label that did arrive: the one AI Act duty that came on time lives in the productGovernance19 min
- 30 July 2026An AI lab’s own models broke into Hugging Face. Read the postmortem.Security13 min
- 28 July 2026Kimi K3: the weights you can hold, and the arithmetic before you doOpen weights21 min
- 25 July 2026Opus 5 at the old price: an unchanged rate card is not an unchanged billModel strategy20 min
- 22 July 2026Tokens to done: the benchmark the labs started competing on in JulyUnit economics14 min
- 16 July 2026Grounding is the product: what GaRAGe shows about enterprise RAGRetrieval18 min
- 8 July 2026An agent is a principal: identity is the control no vendor ships across vendorsSecurity20 min
- 3 July 2026The evaluation bench, in full: how to decide which model runs each taskEvaluation19 min
- 27 June 2026The browser is the API now, and RPA’s risks are back at ten times the scaleAgentic systems19 min
- 25 June 2026Routing by task class: a reference design for the enterprise model routerPlatform design19 min
- 16 June 2026Indic models: the gap between the benchmark and the bazaarSovereign AI19 min
- 13 June 2026The model that switched off: a frontier model is now a dependency on a day’s noticeModel strategy19 min
- 10 June 2026Twenty billion parameters in your pocket: the device is now a tierModel strategy19 min
- 4 June 2026Ten years of talking machines: what the intent era got rightConversational AI19 min
