Google’s new “information agents” inside Search are not just a consumer feature upgrade; they are a product-level commitment to long-running agent workloads—systems that persist, maintain intent, monitor external change, and trigger actions over time. That shift matters because it aligns consumer UX (alerts, trackers, delegated monitoring) with the same operational requirements enterprise agents have been forced to solve first: state, memory, observability, and cost discipline. In parallel, delta-mem’s attempt to add low-parameter working memory and AWS’s AgentCore case study show the stack converging on a pragmatic view of agents: the winners are not the flashiest reasoners, but the systems that can run continuously, predictably, and cheaply enough to be trusted with real workflows. That is the context in which Goldman’s token-growth model reads less like abstract extrapolation and more like an emerging systems reality.
From “Answer Engine” to “Standing Order”: Search as an Agent Runtime
The key change implied by Search-based information agents is temporal. Classical web search is episodic: a query, a response, and the user carries the burden of follow-up. An information agent reverses that burden by turning a user’s intent into an ongoing process that repeatedly checks the world and decides when to notify.
Why persistence is the product, not the model
Search has always been a retrieval system; the novel element is a first-party, user-managed layer of persistence and triggers. Once the user can say “watch this topic and alert me,” the value proposition becomes closer to a standing order than a query.
This has two immediate consequences that are visible in Google’s rollout choices. First, it introduces an operational cadence (background monitoring and push notifications) that is inherently “agentic” even when the actions are narrow. Second, it supports subscription gating (Pro/Ultra), which is a rational monetization move precisely because persistent agents are recurring compute obligations rather than ad-monetizable single interactions.
Consumer agents quietly normalize the enterprise requirement set
Consumer monitoring agents appear lightweight—movie tickets, apartments, product drops—but they force the same questions enterprises ask:
- What is the agent allowed to do, and how does a user inspect or revoke it?
- How does the agent represent state over time (what it has already checked, what counts as “new,” what thresholds trigger an alert)?
- How do you prevent notification spam, drift, and hallucinated “updates”?
The strategic point is that when a platform like Search begins to carry those responsibilities at scale, it normalizes the expectation that agents are accountable processes, not just conversational interfaces.
Memory Is Becoming an Engineering Budget Line, Not a Research Luxury
Delta-mem is notable less for the specific technique than for what it signals: memory is being re-framed as an efficiency problem that can be solved with small, composable modules rather than ever-expanding context windows.
The hidden cost center in long-running agents: repeated context reconstruction
Many deployed agent patterns today rely on one of two expensive crutches:
- stuffing more history into the prompt (larger context windows), or
- repeated retrieval and summarization cycles (RAG-plus-summarize).
Both approaches burn tokens as a tax on continuity. If an agent is meant to run all day—monitoring, checking, updating, re-planning—then “continuity tax” becomes a dominant operating cost.
Delta-mem’s promise (as described) is a route to continuity with minimal parameter add-on and without changing the backbone model, which is operationally attractive: it suggests memory can be layered in as a component in an agent stack, the way caching layers and indices are layered into traditional systems.
Why this matters for reliability, not just cost
In practice, many agent failures aren’t raw reasoning failures; they are continuity failures: forgetting constraints, re-asking solved questions, losing the thread between steps, or oscillating between plans. A working-memory module aimed at stabilizing “what the agent believes it is doing” is therefore also a reliability play. That reliability improvement is the prerequisite for broader automation—because enterprises do not scale systems that require constant human babysitting.
The Enterprise Agent Stack Is Converging on Auditability and Measurable Outcomes
The AWS OPLOG case study is important as a reference architecture because it treats agents as production services with measurable business impact, not demos. The reported outcomes—reduced sales cycles, improved CRM completeness, reduced manual research—are framed as operational metrics, which is how enterprise budgets get allocated.
Multi-agent deployment: independence plus coordination
AWS describes three agents operating independently across fragmented tools (HubSpot, Teams, Databricks). That “independent agents + shared environment” pattern is one of the more durable multi-agent shapes in enterprise because it maps to organizational reality: sales ops, analytics, and customer-facing workflows have different constraints and data surfaces. A single monolithic agent tends to either overreach in permissions or become brittle in scope.
Observability is becoming part of the product surface
Even when not highlighted, the presence of an “AgentCore” style runtime implicitly elevates runtime controls—deployment, evaluation, and monitoring—from afterthoughts to first-class capabilities. This is the enterprise readiness thread that makes agentic automation believable at scale: agents are being packaged in ways that allow operators to see what happened, reproduce it, and change it without rewriting the world.
The connective tissue to consumer Search is that Google is also introducing a management surface (tracked topics in AI Mode history). Different domain, same direction: persistent agents require user/operator-facing control planes.
Token Economics: Why the “Compute Multiplier” Thesis Has a Mechanism Now
Goldman’s model—token consumption multiplying over 2026–2030 alongside steep per-token cost declines—rests on a simple observation: agents turn one intent into many steps, and each step consumes inference.
Steps, not chats, are the driver
Traditional chat usage is often bounded by a short session. Agent usage is bounded by task completion and ongoing monitoring cycles. A single “track apartments with these amenities” standing order can generate repeated web checks, extraction, de-duplication, threshold evaluation, and notification formatting. None of this is exotic; it’s just iterative execution.
The point is that consumer information agents and enterprise BI agents share this characteristic: they are programs that run repeatedly, not conversations that end.
Cost declines unlock workload migration, not just margin expansion
Goldman’s cited annual inference cost declines (60%–70%) matter because they shift which tasks are economically automatable. As costs drop, systems can afford redundancy (verification calls, multi-pass extraction) and higher-frequency polling—both of which improve quality and timeliness. That creates a feedback loop: better quality increases adoption; adoption increases total tokens even if unit costs fall.
Today’s evidence supports this mechanism: we already see subscription gating for persistent consumer agents, and we see enterprise case studies reporting hard operational gains. Those are two different budget sources paying for the same underlying phenomenon—more long-running inference.
What This Means for the Agentic Economy
The agentic economy depends less on any single breakthrough model and more on whether agents can be treated as auditable, continuously running services with acceptable unit economics. The stories here collectively indicate that this is becoming true in three reinforcing ways.
First, Google is productizing persistence for consumers through Search-based information agents. That normalizes delegated monitoring as a mainstream behavior and shifts agent value from “better answers” to “managed intent over time,” which is a service category that can sustain recurring revenue.
Second, the technical stack is beginning to price continuity explicitly. Memory approaches like delta-mem reflect a push to reduce the continuity tax that makes long workflows expensive and flaky. Lower-cost, more stable memory is not a nice-to-have; it is what allows agents to run longer without either ballooning context or repeatedly reconstructing state.
Third, enterprise deployments are moving from pilots to operational systems with measurable outcomes, as reflected in AWS’s reference architecture and reported business metrics. When agent runtimes come with orchestration and control surfaces—mirrored, in simpler form, by consumer management UIs—agents become governable. Governability is what allows agents to transact on behalf of organizations, touch sensitive systems, and execute workflows without constant human intervention.
Put together, these trends give Goldman’s token-boom thesis a concrete underpinning: not “more AI,” but more persistent, multi-step processes embedded into platforms and business systems. The agentic economy will scale to the extent that persistence, memory, and observability become defaults rather than bespoke engineering—because that is what turns agents from impressive interactions into dependable economic actors.
Sources
https://finance.yahoo.com/sectors/technology/article/google-unveils-biggest-update-to-search-in-25-years-including-ai-agents-174500670.html https://techcrunch.com/2026/05/19/how-to-use-googles-new-ai-agents-to-go-beyond-your-standard-searches/ https://venturebeat.com/orchestration/a-0-12-parameter-add-on-gives-ai-agents-the-working-memory-rag-cant https://www.goldmansachs.com/insights/articles/ai-agents-forecast-to-boost-tech-cash-flow-as-usage-soars https://developers.googleblog.com/adk-kotlin-android-building-ai-agents/ https://aws.amazon.com/blogs/machine-learning/build-ai-agents-for-business-intelligence-with-amazon-bedrock-agentcore/