A Leading Indicator for the Agentic Software Rotation
25 Names Across 7 Sectors — Structural Leaders, Bellwethers, and Oversold High-Beta Plays
Monitored Through the Temperature Signal Framework and Relative Strength Analysis
March 2026 · For Qualified & Institutional Investors Only
Standard equity indices are archaeological instruments. They catalogue what has already happened — which companies grew large enough to enter a committee-selected basket, weighted by a market capitalisation that reflects yesterday's consensus. The S&P 500 tells you where capital has accumulated. It tells you nothing about where capital is migrating, where supply chains are breaking, or where capability transitions are creating the conditions for the next decade of value creation.
Closelook's functional indices are designed to solve a different problem. Rather than reflecting the market as it was, they attempt to map the market as it is becoming. Each index is constructed from a structural thesis — a specific claim about where the economy's weight-bearing architecture is shifting — and the constituents are selected not by market capitalisation or committee consensus, but by their functional role in that structural transition.
The Closelook index family divides into two categories. Strategic indices operate on multi-year horizons and track the slow, powerful currents of structural positioning. The Rubin Build-Out 100 maps the physical infrastructure constraint layer across 18 sectors and 100 companies — from EUV lithography to advanced packaging to power delivery. The Euro-AI Sovereign 50 tracks European-listed companies whose revenue intensity in the AI value chain is systematically underpriced by a market fixated on US mega-caps. The HALO Functional Index — Hardware, Assets, Logistics & Operations — captures what we call "Smart Atoms": the physical-world companies that benefit from AI-driven automation without being AI companies themselves.
Tactical indices operate on shorter horizons — typically six months — and are regime-sensitive. They answer not "where is the structural value?" but "when does the market start pricing it?" The Agentic Winners 25 is the first tactical index in this family. Its constituents are not selected purely for structural positioning — some are there because they are the deepest workflow platforms, but others are included as bellwethers, as the most oversold names in the sector, or as the stocks that will move first and hardest when sentiment shifts from "very bad" to "not so bad." The basket is designed as a seismograph: a leading indicator for whether the agentic software sector is beginning to rotate — the way semiconductors did in early 2024 before the consensus caught on.
The key distinction: Strategic indices are instruments for portfolio construction. Tactical indices are instruments for signal detection. If you wanted the five best structural agentic plays for a five-year hold, you would build a different list. The AW25 is not that list. It is a 25-name basket calibrated to capture the earliest observable evidence of sector rotation — through a mix of structural leaders, beaten-down bellwethers, and high-beta names whose price action will tell you the story before the fundamentals confirm it. You own the strategic indices through cycles. You watch the tactical indices for the moment the cycle turns.
The cost structure of running AI agents has been collapsing in a way that most investors have not yet internalised. In March 2026, Google published TurboQuant — a technique that achieves six-fold KV cache compression without any retraining. By itself, this is an engineering paper. In combination with speculative decoding, sparse mixture-of-experts routing, aggressive quantisation, context caching, and distillation, the effect is multiplicative. A workflow that cost three to five dollars per execution eighteen months ago now costs cents.
This is not a theoretical projection. It is already visible in production deployments. An agentic workflow that reads a contract, extracts key terms, cross-references against a compliance database, drafts a summary, and routes it for human approval — a workflow that required a senior paralegal and two hours in 2024 — can now be executed by a chain of specialised agents for less than the cost of a cup of coffee. The economic threshold for agent deployment has crossed from "impressive demo" to "obvious replacement" in the span of two quarters.
The DeepSeek moment in late 2025 was about training efficiency. The market correctly identified the Jevons Paradox implications: cheaper training means more models, more experimentation, more demand for compute, not less. TurboQuant and its successors represent the same dynamic on the inference side — and inference is where the recurring revenue lives. Training is a one-time capital expenditure. Inference is metered consumption. Every agent call, every tool use, every chain-of-thought execution generates an inference transaction. When the unit cost of that transaction drops by an order of magnitude, the addressable market expands by multiples.
Google's Universal Commerce Protocol, announced in January 2026, provides the connective tissue. UCP establishes a common language for AI agents to negotiate purchases, compare products, verify inventory, and execute transactions on behalf of users. Combined with compressed inference models that can run on consumer hardware — a 30-billion-parameter model operating within the KV cache budget of an M5 MacBook — the agent economy is no longer confined to cloud-native enterprises. It extends to every small business owner, every freelancer, every household with a capable laptop.
The agentic stack is crystallising into a seven-layer structure, from silicon (NPU inference chips) through model runtime, orchestration frameworks, tool integration, enterprise workflow platforms, and finally the application interface layer. Below all of this sits a backdoor exposure layer — the companies whose revenue models are threatened by agents performing the work their products were designed to support.
The critical insight for investors is that the demand side of AI is now forming — and it is potentially larger than the supply side. The Rubin Build-Out thesis captured the supply side: the physical infrastructure required to train and deploy frontier models. The Agentic Winners thesis captures the demand side: the companies through which agent-driven economic activity will flow. If the supply side was the construction of the railroad, the demand side is the commerce that rides the rails.
Phase 1 (Now — H2 2026): Copilot features embedded in existing platforms. Agent-assisted workflows with human-in-the-loop. Cost savings visible in operational metrics but not yet in revenue models.
Phase 2 (2027–2029): Autonomous agent workflows. Multi-agent chains executing complex business processes end-to-end. Platform companies become agent operating systems. Revenue model transition from seat-based to usage-based.
Phase 3 (2029+): Agent-native businesses. Companies built from inception on autonomous agent architectures. The distinction between "software company" and "agent platform" dissolves.
In the Rubin Build-Out universe, stock selection is relatively straightforward. ASML makes extreme ultraviolet lithography machines. Advantest makes semiconductor test equipment. Micron manufactures memory chips. The revenue streams are legible, the competitive moats are visible, and the supply chain dependencies are mappable. A lithography company is a lithography company — there is no ambiguity about what you are buying.
Agentic software is categorically different. Every company in this universe is a hybrid. SAP generates the vast majority of its revenue from legacy ERP licence and maintenance contracts — a business model that has existed since the 1990s. Simultaneously, SAP is building an agentic layer — Joule — that can autonomously execute procurement workflows, generate financial reports, and manage human capital processes. The market cannot easily separate "old SAP" from "agentic SAP." This creates mispricing in both directions: companies with genuine agent platform potential trade at a discount because their legacy revenue depresses growth multiples, while companies with cosmetic AI features trade at a premium because the market cannot yet distinguish decoration from architecture.
The transition from Phase 2 to Phase 3 is the trade. We are currently in the indiscriminate fear phase, where the market treats enterprise software as a monolith — uniformly threatened by AI disruption, uniformly deserving of multiple compression. This is the same pattern that played out in semiconductors during the memory downcycle of 2022–2023: the market sold everything, made no distinctions between commodity DRAM and structural HBM demand, and then repriced violently upward when the differentiation became undeniable.
The question that defines Phase 3 is not "does this company survive AI?" but "does this company become the operating layer for agents?" That question determines the structural winners. But for the AW25 — a tactical instrument, not a conviction portfolio — the question is different: "which basket of stocks will move first and most when the market starts asking that question?"
Five structural attributes separate the companies that become agent infrastructure from those that become agent casualties. First, systems of record combined with data gravity — the platform holds the authoritative dataset for a business function, and agents must interface with it because the data cannot be replicated elsewhere. Second, mission-critical workflow positioning — the platform sits in the execution path of processes that cannot tolerate failure, latency, or hallucination. Third, regulatory and compliance infrastructure — the platform enforces audit trails, access controls, and regulatory reporting that agents cannot bypass. Fourth, agent deployment capability — the platform is not merely using AI internally but is becoming the surface on which external agents operate. Fifth, network effects that strengthen with agent volume — more agents using the platform generates more data, better models, and deeper integration, creating a flywheel that competitors cannot easily replicate.
Twenty-five companies across seven sectors. The selection blends three types of constituent. Structural leaders — platforms with genuine agent operating layer potential, deep workflow embedding, and data gravity that agents cannot bypass. Bellwethers — large, liquid names like Microsoft, Salesforce, and Intuit whose price action telegraphs sector sentiment before mid-caps move. And high-beta oversold names — stocks beaten down hardest during the Phase 2 fear cycle, where the swing from "very bad" to "not so bad" produces the largest and earliest price moves. The basket is not a conviction portfolio. It is a sensor array. The sector structure reflects function in the agentic value chain, not industry classification.
The mega-cap platforms through which agentic adoption necessarily flows. Microsoft's Copilot ecosystem embeds agents into the Office suite used by over a billion knowledge workers. Google's Gemini infrastructure powers both the model layer and the commercial agent protocol (UCP). Apple controls the on-device inference environment — the hardware runtime on which personal agents will execute. These are not software companies in the traditional sense; they are the plumbing of the agent economy. They occupy a separate sector because their beta characteristics, valuation frameworks, and capital flow dynamics are fundamentally different from mid-cap enterprise software. When the market rotates into "AI software," these move first and most; when it rotates out, they provide relative safety. MSFT sits at critical multi-year support at $356 — a level that has held through three distinct selloffs, suggesting institutional accumulation at scale.
The deepest workflow embedding in enterprise. SAP's ERP backbone processes 77% of global transaction revenue — agents executing procurement, finance, or supply chain workflows must interact with SAP regardless of which model powers them. ServiceNow owns IT service management and is extending into enterprise-wide workflow orchestration; its agent platform is evolving from copilot to autonomous executor. Salesforce's CRM dataset — the record of every customer interaction — is the training ground for sales and service agents. Workday holds the human capital management record: payroll, benefits, compliance, org structure. Each of these platforms has decades of legacy revenue providing a valuation floor, while agentic capabilities create upside optionality that the market, stuck in Phase 2 fear, is currently pricing at close to zero.
These companies win on rising agent transaction volume regardless of which model, which orchestration framework, or which application prevails. They are the Advantest equivalents of the agent era — the pick-and-shovel plays whose revenue scales with activity, not with any single platform's success. Cloudflare delivers edge compute and secures agent-to-API traffic. Datadog observes agent behaviour, monitors latency, and traces failures across distributed agent chains. CrowdStrike secures the endpoints and cloud workloads that agents access — and as agent surface area expands, so does the attack surface. Twilio provides the communication APIs through which agents interact with customers via voice, SMS, and messaging. Okta manages identity and authorisation — every agent action that touches user data must pass through an identity verification layer. The toll booth thesis is model-agnostic and platform-agnostic: whoever wins the agent platform war, the toll collectors get paid on every transaction.
The infrastructure for deploying inference at the edge. DigitalOcean serves the developer-cloud segment where agent hosting for startups and SMBs will concentrate. Fastly provides sub-50ms edge latency — critical for real-time agent interactions. Qualcomm's Snapdragon NPU silicon is the on-device inference engine, enabling agents to run locally without cloud round-trips. As inference economics improve and models shrink, the edge becomes the primary execution environment.
Agents need structured data access, real-time state management, and vector search — the retrieval-augmented generation (RAG) infrastructure. Snowflake provides the governed data warehouse agents query for analytical context. MongoDB offers the flexible document store for agent state and configuration. Elastic is the search and retrieval backbone — its vector search capabilities power the RAG pipelines that ground agent responses in enterprise data. These are the picks and shovels of the data layer.
Vertical software commands the highest margins, the deepest moats, and three-to-five times the retention rates of horizontal platforms. The reason is structural: domain knowledge, regulatory compliance requirements, and industry-specific data lock-in create barriers that no general-purpose agent can bypass. Veeva Systems owns the pharma and life sciences workflow — clinical trials, regulatory submissions, and drug safety data live in Veeva's vault, and any agent operating in that domain must interface with it. Intuit holds the SMB financial record: tax, accounting, and payroll for millions of small businesses, wrapped in compliance infrastructure that took decades to build. PTC's industrial IoT and PLM platforms manage the digital twins and product lifecycles that manufacturing agents will need to read and write. CoStar Group controls the commercial real estate data monopoly — property records, transaction history, and market analytics that no agent can replicate from public sources. Samsara's IoT platform connects the physical fleet, warehouse, and equipment assets that operational agents need to monitor and manage. These are not software companies that happen to serve a vertical. They are the system of record for their respective industries.
Two companies that are building the agent-native infrastructure from the ground up, rather than retrofitting existing products. Shopify is the commerce operating system for the agent economy — its integration with Google's UCP positions it as the platform through which AI shopping agents will browse, compare, and purchase on behalf of consumers. When agents become the primary shopping interface, Shopify merchants are already connected to the protocol. Palantir occupies a unique position as the enterprise AI orchestration layer — its platforms (Foundry, AIP) are designed to integrate data from disparate silos, deploy AI models against operational decisions, and maintain the governance and audit infrastructure that enterprise agent deployment requires. PLTR shows persistent relative strength, holding an ascending trendline through market-wide selloffs — a technical signature of institutional conviction in the agentic thesis.
| Sector | Ticker | Company | Role | Tag |
|---|---|---|---|---|
| S1 | MSFT | Microsoft | Copilot ecosystem, Azure agent infra | Gateway |
| S1 | GOOGL | Alphabet | Gemini models, UCP, cloud agent runtime | Gateway |
| S1 | AAPL | Apple | On-device inference, personal agent runtime | Gateway |
| S2 | SAP | SAP SE | ERP backbone, Joule agent layer | Platform |
| S2 | NOW | ServiceNow | IT workflow orchestration, agent platform | Platform |
| S2 | CRM | Salesforce | CRM data record, Agentforce | Platform |
| S2 | WDAY | Workday | HCM record, payroll & compliance | Platform |
| S3 | NET | Cloudflare | Edge delivery, agent-API security | Toll Booth |
| S3 | DDOG | Datadog | Agent observability, trace & monitor | Toll Booth |
| S3 | CRWD | CrowdStrike | Endpoint & cloud security | Toll Booth |
| S3 | TWLO | Twilio | Communication APIs for agent interaction | Toll Booth |
| S3 | OKTA | Okta | Identity & authorisation for agent actions | Toll Booth |
| S4 | DOCN | DigitalOcean | Developer-cloud, agent hosting | Neocloud |
| S4 | FSLY | Fastly | Sub-50ms edge latency | Neocloud |
| S4 | QCOM | Qualcomm | Snapdragon NPU, on-device inference | Neocloud |
| S5 | SNOW | Snowflake | Governed data warehouse for agents | Data |
| S5 | MDB | MongoDB | Document store, agent state | Data |
| S5 | ESTC | Elastic | Vector search, RAG infrastructure | Data |
| S6 | VEEV | Veeva Systems | Pharma & life sciences record | Vertical |
| S6 | INTU | Intuit | SMB finance, tax & payroll | Vertical |
| S6 | PTC | PTC | Industrial IoT, PLM, digital twin | Vertical |
| S6 | CSGP | CoStar Group | Commercial real estate data monopoly | Vertical |
| S6 | IOT | Samsara | Physical operations IoT | Vertical |
| S7 | SHOP | Shopify | Agent commerce layer + UCP | Commerce |
| S7 | PLTR | Palantir | Enterprise AI orchestration | Orchestration |
By the time the consensus recognises "agentic software is the new narrative" — by the time the sell-side upgrades cascade, the conference panel invitations go out, and the CNBC anchors start saying the word "agent" in every segment — the stocks have already moved. This is not cynicism; it is the observable mechanics of institutional accumulation. Large funds cannot buy quickly without moving price. They accumulate over weeks, in the noise, on down days, at technical support levels. The chart tells you before the analyst upgrade does.
The Temperature signal framework is designed to detect this accumulation before it becomes visible in headlines. It synthesises five dimensions — position in the moving average stack, momentum acceleration, volume conviction, volatility state, and cross-instrument regime fragility — into a single reading from 0 to 100 per instrument. But the power is not in the individual reading. It is in the pattern across the basket.
When price tests the 200-day moving average, volume tells the story. A high-volume bounce means strategic buyers are present — institutions defending a level. A low-volume drift through the 200-DMA means they have not yet arrived. This single observation separates "support is holding" from "support is an illusion."
On-Balance Volume measures cumulative buying and selling pressure. When price makes lower lows but OBV makes higher lows, accumulation is occurring beneath the surface — invisible in price, visible in flow. This divergence is the earliest quantitative signal of institutional positioning.
Weekly RSI divergence — momentum exhausting to the downside while price continues to fall — indicates that selling pressure is weakening even as the narrative remains bearish. This typically precedes a base formation by two to four weeks.
When the 50-day and 200-day moving averages converge and price oscillates in a narrowing range between them, the stock is coiling. Trend followers are selling (below the 50d), institutions are buying (above the 200d). The resolution — in either direction — is typically violent and tradeable.
The most powerful signal is not what happens on green days — it is what happens on red days. When the Nasdaq drops 2%, does the AW25 basket drop 3% (pure beta, no differentiation) or 1.5% (institutional hands refusing to sell)? Beta-adjusted drawdown on selloff days is the earliest observable signal of a rotation. It appears before relative performance metrics, before fund flow data, and before sector ETF creation.
Recovery day behaviour adds a second layer. After a broad market selloff, does the AW25 basket lead or lag the index on the bounce? If mid-cap software starts consistently outperforming the MAG7 on bounce days — even marginally — the rotation has begun. Money is moving down the capitalisation curve into the structural beneficiaries.
No single indicator triggers a position. The entry signal requires a conjunction of conditions: the Rubin Build-Out thesis pauses or consolidates — infrastructure capex cycle reaching maturity — while software names simultaneously show relative outperformance on selloff days. Double bottoms form on long-term charts. Three or four names in the AW25 basket show the Temperature pattern simultaneously — high-volume 200-DMA defence, OBV divergence, weekly RSI divergence, 50/200 compression. A cluster, not an outlier. That is the sector entry signal.
The playbook has a precedent. SK Hynix and Micron in early 2024. Both stocks were deep in a cyclical memory washout. The consensus narrative was "memory oversupply, structurally impaired." Institutional buyers quietly accumulated through the entire downturn. By the time the market recognised the HBM structural story — that high-bandwidth memory was not commodity DRAM but a capacity-constrained product with eighteen-month lead times — both stocks had already doubled from their lows.
The agentic software trade follows the same script. The consensus says "AI kills SaaS." The reality is more nuanced: AI kills some SaaS and crowns the survivors as essential infrastructure. The distinction is not yet priced. The Temperature signal is designed to tell us when the market begins to make that distinction.
The Agentic Winners 25 is a template. The construction method — identify a structural thesis, build a focused watchlist, monitor with Temperature signals, trigger on relative strength — is replicable across any macro rotation that meets two criteria: the structural case is sound, and the market has not yet priced it.
Each tactical index has a shelf life. It fires or it does not within approximately six months, at which point it is either promoted to a strategic index (if the thesis proves durable and the constituents stabilise) or retired (if the macro conditions shift and the trade window closes). This is not index investing in the traditional sense — it is systematic watchlist management with quantitative entry triggers.
| Tactical Index | Thesis | Status | Horizon |
|---|---|---|---|
| Agentic Winners 25 | Enterprise software survivors become agent operating layer | Active | H1–H2 2026 |
| De-dollarization Basket | Gold + BTC + non-US equities benefit from dollar regime shift | Monitoring | TBD |
| European Re-rating Basket | Euro-AI divergence triggers institutional reallocation to EU | Monitoring | TBD |
| Energy / Grid Infrastructure | AI power demand drives grid modernisation capex | Research | TBD |
The Temperature signal framework — described in detail in a companion specification — provides the common monitoring language across all tactical indices. The same five dimensions (Position, Momentum, Volume, Volatility, Fragility) apply to any asset class. The regime classifier — which pattern-matches across eight macro instruments to identify risk-on, risk-off, rotation, and liquidity conditions — serves as the macro overlay for all tactical positioning. When the regime says "risk-off," no tactical index fires regardless of individual Temperature readings. The macro environment must be permissive before the tactical signal matters.
Constituents are selected across three overlapping criteria, reflecting the index's role as a leading indicator rather than a structural conviction portfolio. Structural relevance: the company plays a meaningful role in the agentic value chain — as a platform, infrastructure provider, or data layer. This does not require satisfying all five winner characteristics; a single strong structural attribute (e.g. system of record, toll booth positioning) suffices. Bellwether signal value: the company is large and liquid enough that its price action reliably leads sector sentiment — institutional flows into these names precede and predict flows into the broader agentic software universe. Oversold beta: the company has been disproportionately punished during the Phase 2 fear cycle, creating conditions where the swing from extreme pessimism to moderate reassessment produces the earliest and largest price moves in the basket. Most constituents satisfy two or three of these criteria simultaneously. Sector assignment reflects functional role in the agentic stack, not GICS industry classification.
The AW25 Composite Index uses equal weighting at the sector level. Each of the seven sectors receives an equal share of the index weight, and within each sector, constituents are equally weighted. This ensures that the index reflects the thesis across the full agentic stack rather than being dominated by MAG7 market capitalisation. Sub-indices S1 through S7 are published separately for sector-level analysis. The base value is 1,000, set from March 31, 2026.
Each constituent receives a daily Temperature score from 0 to 100, computed across five dimensions: Position (30 points) — where price sits in the moving average stack and relative to Bollinger Bands; Momentum (25 points) — rate of change velocity and acceleration, RSI divergence; Volume Conviction (20 points) — OBV trend alignment, accumulation-distribution, volume at MA confirmation; Volatility State (15 points) — ATR normalised percentile, Bollinger width, ADX trend strength; and Regime Fragility (10 points) — absorption ratio derived from the cross-instrument correlation matrix. Full scoring methodology is documented in the Temperature specification.
Daily: OHLCV data pull for all 25 constituents at 23:00 UTC, Temperature score computation, R2 JSON output for site embeds. Weekly: Temperature scan summary published to C+ subscribers — narrative assessment of which names are heating, cooling, or coiling, with regime context. Monthly: constituent review — any name that has structurally deteriorated (lost a key competitive position, been acquired, or failed to maintain relevance to the agentic thesis) is flagged for potential replacement. Data sources: EOD Historical Data for daily OHLCV, Barchart Premier for manual chart validation.
The AW25 rebalances quarterly. Constituents are reviewed against their selection rationale — whether structural relevance, bellwether signal value, or oversold beta sensitivity. A name is removed if its role in the basket has fundamentally changed: a bellwether that is no longer liquid enough to lead, a structural play that has lost its competitive position, or a formerly oversold name that has re-rated to fair value and no longer offers early-signal sensitivity. Price performance alone is not a removal criterion — the index does not chase momentum. If a constituent is acquired or pivots away from the agentic software universe, it is replaced at the next quarterly review.
The supply side of AI has been the consensus trade for two years. Infrastructure capex — GPUs, memory, networking, power — has consumed the market's attention and capital. The Rubin Build-Out thesis captured that trade.
The demand side is forming now. As inference costs collapse, agent frameworks mature, and commerce protocols standardise, the question shifts from "who builds the railroad" to "who operates the stations." Somewhere in the current wreckage of enterprise software valuations are the companies that will become the operating layer for autonomous AI workflows. Identifying them with certainty today is premature. But detecting when the market starts to differentiate — when capital begins flowing selectively rather than fleeing indiscriminately — that is observable.
The AW25 basket is designed to capture that signal. It mixes structural leaders with bellwethers and beaten-down high-beta names precisely because the earliest evidence of rotation appears in all three simultaneously: institutions defend support on the bellwethers, the most oversold names bounce hardest on relief rallies, and the structural leaders start outperforming on selloff days. When these patterns align across the basket — not one stock, but a cluster — the sector is turning.
The Agentic Winners 25 is not a prediction that these 25 stocks will rise. It is not the list of the best structural agentic plays for a five-year hold. It is a seismograph — a structured framework for detecting the moment when the market begins to differentiate winners from casualties in enterprise software.
This strategy paper is provided for informational and analytical purposes only and does not constitute investment advice, a solicitation to buy or sell any security, or an offer of any financial product. The Agentic Winners 25 is a research index and does not represent a managed portfolio or investable product. Temperature scores, relative strength signals, and regime classifications are analytical tools, not trading recommendations. Simulated index returns are based on constituent performance and equal-weight methodology; they do not reflect actual tracked index performance. Past performance, whether actual or simulated, is not indicative of future results. All data sourced from public filings, EOD Historical Data, and Closelooknet analysis as of March 2026. Investors should conduct their own independent due diligence and consult with qualified advisors before making any investment decisions.
