How operators decide where AI should change coordination, judgment, learning speed, and commercial leverage.
Research synthesis · July 2026
Agents need a memory architecture.
A field-grounded research synthesis for founders, consultants, and C-suite operators. Production agents store experience in prompt rules and model weights that drift apart unless someone designs ownership and reconciliation on purpose.
People allocating scarce resources under noisy market pressure.
Each memo ends with a call you can act on this week.
Verticals
Three beats run alongside the main AI strategy feed.
Build logs, failure modes, and operating decisions from running a small fleet of AI agents day to day.
Payments rails, bank-fintech partnerships, stablecoins, and licensing — who carries the risk when a new rail routes around an old one.
Founders, capital, and supply chains at the US–ASEAN–China intersection, treated country by country rather than as one region.
Browse by topic
Filter the memo index by topic, or show everything.
Showing all memos
Latest Memo
Agents Need a Memory Architecture
Field synthesis from 1,990 public discussions. Production agents hit a memory fork between prompt rules and model weights — three episodes on retention pressure, parallel stores, and coherence stress, before any checklist would help.
Reader decision
Which memory stores does the agent write, and who owns reconciliation?
When AI Disclosure Is Due
EU AI Act Article 50 transparency duties apply on 2 August 2026. Commission guidelines published 20 July turn disclosure into a workflow deadline — interactive AI, synthetic content, and who still owns residual risk.
Reader decision
Which customer-facing AI surfaces still hide the machine?
Where AI Coverage Still Fails
Generative AI exclusions are turning residual risk transfer into a management problem. Operators need exposure classes, policy lines, residual owners, control evidence, and stop rules before autonomy scales.
Reader decision
Which AI acts remain uninsured residual risk?
What Browser Agents Can Touch
Agentic browsers turn a reading surface into an execution surface that carries employee credentials. Operators need clear limits for session privilege, instruction channels, tool reach, data movement, work classes, and incident paths.
Reader decision
Which browser agents may act with employee credentials?
Who Can Use Your Source Material
AI search is unbundling the old crawler bargain into access, citation, answer, training, payment, and relationship rights. Operators need a source-rights map before publishing strategy becomes accidental licensing by default.
Reader decision
Which source rights are you granting by default?
Who Is Accountable When AI Acts
AI accountability is reaching executives before visibility, spend discipline, incident response, and stop rules have caught up. Operators need a accountability table before agents harden into infrastructure.
Reader decision
Where has AI accountability already outrun control?
What You Keep After AI Search
AI search is weakening the old traffic bargain. Operators need a clear view of crawl access, answer presence, referral quality, owned relationships, conversion rights, and proof loops.
Reader decision
What does the company keep when AI search consumes, cites, or summarizes its public work?
The AI Slop Audit
AI adoption is broad enough that low-grade AI output has become a management problem. Operators need to audit accuracy, specificity, judgment, review burden, and learning before volume quietly becomes the strategy.
Reader decision
Where is AI increasing accepted work, and where is it increasing review burden?
What Agents Remember
AI assistants are starting to carry context across sessions. Operators need a clear list of preference memory, role memory, project memory, sensitive memory, and official institutional memory.
Reader decision
Which AI memories should persist, and under whose authority?
Which Models for Which Work
AI buying is becoming a portfolio problem. Operators need a clear view of frontier APIs, open-weight models, vendor bundles, local hosting, specialist models, and work that should stay off-limits.
Reader decision
Which workflows deserve which model class?
Why AI Productivity Does Not Show Up
AI investment and awareness are running ahead of visible productivity proof. Operators need a clear view of adoption, daily use, workflow redesign, measurement, learning, and capital discipline before the gap widens.
Reader decision
Which workflow has earned an AI productivity proof?
When Distribution Stops Belonging to You
AI search, workplace agents, and browser agents are turning distribution into an ownership problem. Operators need to know where buyer intent is interpreted, who frames the options, and which action path the company still owns.
Reader decision
Which boundary owns the next customer action?
Who the Agent Is Acting As
Browser agents and tool-using AI systems turn identity into a management control. Operators need to know what acted, who granted authority, which tools were in scope, and how access can be revoked.
Reader decision
Where does borrowed identity become too risky for an agent workflow?
What an AI Disclosure Record Must Contain
AI transparency is becoming a workflow problem. Operators need to know what was generated, where it travelled, who approved it, and when disclosure or provenance should stop a piece from going out.
Reader decision
Where does disclosure become material enough to require a sheet, not just a label?
What Evaluation Still Misses
Benchmarks and demos are not enough. Operators need a live evidence loop that proves task fit, workflow fit, failure response, cost burden, and learning schedule before AI delegation scales further.
Reader decision
What evidence would force an AI pilot to pause, shrink, or stop?
What Agents Are Allowed To Do
Agentic AI turns access into strategy. Operators need to decide what an agent may read, reason about, write, spend, message, and change — before the pilot quietly becomes production authority.
Reader decision
Which permission line should remain gated before agent autonomy expands?
How Often AI Work Should Run
AI adoption is moving from occasional prompting into recurring production. Operators need to manage the artifact, the rhythm, the human handoff, the review burden, and the job redesign that follows.
Reader decision
What recurring workflow should get a work schedule before AI delegation expands?
How Much Inference a Workflow Deserves
Inference is becoming a management budget. Executives need to decide which workflows deserve expensive model tiers, long context, agent loops, and elevated system permissions.
Reader decision
Where should intelligence be spent, capped, downshifted, or stopped?
Proof That AI Changed the Work
AI benchmarks keep improving, but executives still need proof that capability turned into task output, workflow change, economic gain, and a control burden they can live with.
Reader decision
Which proof metric must move before an AI pilot becomes an operating commitment?
Where Founders Should Publish
Treat founder publishing as a four-column map — shelf, relationship, conversion, proof — so channel fashion does not masquerade as strategy. Tools are examples; the columns are what actually matter.
Reader decision
Which column is weak before you add another channel?
Treat Each ASEAN Country as a Separate Market Test
ASEAN is writing better digital rules. Founders should still treat Southeast Asia as a portfolio of country-level operating tests — not as proof that one regional market has already arrived.
Reader decision
Which first ASEAN country best tests the buyer, trust bridge, compliance burden, payment stack, and reusable operating model?
The Stablecoin Distribution Test
Stablecoins are moving from crypto infrastructure into payment operations. For a small digital business, the question is whether the rail unlocks buyers and settlement speed at an operating cost the founder can actually carry.
Reader decision
Should a founder add stablecoin payment access now, wait for a provider-mediated checkout, or keep the first conversion channel on a familiar platform?
Evidence an AI Project Must Produce
AI procurement is moving from demo judgment to evidence judgment. Before buying an agent, model, copilot, or automation layer, executives should know which claims have proof, which risks are owned, and which failures would force a rollback rather than a shrug.
Reader decision
What vendor evidence and buyer-owned workflow facts must be clear before an AI system moves from pilot to production?
When AI Pays for Itself
The AI capex question is not whether the boom is real or fake. It is who owns the payback risk when model ambition is financed through chips, data centers, leases, guarantees, debt, and customer prepayments — and who is left holding it when the story cools.
When Every Team Asks the Same Model the Same Question
When every team can ask similar models similar questions, the scarce asset is no longer analysis. It is independent judgment about which answers deserve trust, which to ignore, and which to deliberately oppose.
Is This AI Quality or Just Theme Exposure?
The harder AI question is no longer whether a company can spend. It is whether that spending builds a durable business advantage — or only buys exposure to a theme everyone else can already see.
Who Can Block a Data Center
The next AI bottleneck may not be the model. It may be the right to draw power, use water, raise capital, win permits — and keep local voters from deciding the buildout is someone else's upside, funded by their bills.
Is This AI Strategy Real?
Fake AI strategy is not the absence of demos. It is demos without a credible theory of workflow change, accountability, evidence, management, and opportunity cost.
AI Needs a Management Layer
Enterprise AI is being folded into managed work, agent registries, observability logs, content provenance, licensing, and transparency rules. The operating question is whether companies can manage delegated AI work before the market forces the issue.
The AI Vendor's Warning Label
Agent-security guidance is becoming the AI industry's shared-responsibility model — a practical control manual today, and potentially an accountability record after the first serious incident.
The AI Pilot Trap
Companies are not short of AI use cases. They are short of managerial judgment about where autonomy should enter the business — and where it should stop.
Decide the Rules Before Buying Tools
Founders do not have an AI tooling problem first. They have a rules problem — a compact set of operating beliefs for deciding what should be automated, assisted, or left alone.