Welcome back to Venture Logbook. This is our first edition of 2026.
To kick off the new year, we're looking at three major reports on what investors and startups are betting on in 2026: a16z Big Ideas 2026, YC's 2025 Fall RFS, and Sequoia's AI in 2026: A Tale of Two AIs.
Each offers a different lens on where the market is heading and what kinds of teams are getting funded. I've picked the themes that caught my attention. Let's break it down.
Multi-Agent
Sequoia describes the market shift from reasoning toward "agentic applications," highlighting that multi-agent systems are starting to go mainstream—which we covered in the last episode.
a16z emphasizes orchestration and coordination as critical for enterprise adoption. YC calls out Infrastructure for Multi-Agent Systems:
AI agents are evolving from single-threaded loops into distributed workflows that fan out many sub-agent calls in a single run. These systems are difficult to build. They require solving traditional distributed systems problems to ensure high throughput and reliability while controlling costs.
The Numbers
- North America's Multi-Agent System market hit $2.38B in 2025 and is projected to grow at 45.1% CAGR
- Dimension Market Research forecasts the global MAS market reaching $184.8B by 2034 (CAGR 45.5%)
- MAS is primarily deployed as software platforms/systems, with Market.us reporting 72.1% of 2024 deployments were cloud-based
My Lens
The power of running multiple agents in parallel is straightforward. You can decompose tasks into subtasks that execute simultaneously, expanding your total budget for reasoning tokens and tool calls. But the difficulty scales with that power: coordination overhead and resource contention grow just as fast.
When agents collaborate, dependencies introduce latency, higher resource consumption, and longer wait times. Beyond inference itself, you need lean workflow design to minimize these bottlenecks. The key to stable multi-agent systems comes down to three things: inference latency, orchestration, and shared state/memory.
Today's multi-agent challenges include:
- Context can't be effectively shared. Sub-agents make conflicting decisions, and integration feels like merging code without communication. Concurrent writes to shared data can cause conflicts and inconsistencies, requiring version control and event sourcing.
- Complex message passing, sequencing, and alignment rules become exponentially harder to debug as agent count increases. Add in resource scheduling and inference queues, and throughput drops as agents compete for the same resources.
This reminds me of when deep research launched. It was arguably the first mainstream multi-agent application. Every service provider rushed to release their own deep search feature to showcase their multi-agent capabilities. I was a heavy user back then, but now I only use it a few times a month.
The strongest use case for multi-agent systems remains coding, especially when paired with cloud-based background agents. It's a path toward 24/7 workers. The economics, however, don't add up. Fitting that kind of compute into a $20/month plan is unsustainable, which explains why Anthropic introduced rate limits and raised pricing to $200/month.
Multi-agent systems essentially combine the workflows of multiple people. They're more flexible than fixed workflows like n8n.
However, is a flat monthly fee really the right business model for multi-agent systems?
AI Native Enterprise Software
YC made this one of their Requests for Startups, drawing parallels to how Salesforce and ServiceNow rose during the cloud wave. They suggest the next generation of AI-native enterprise giants could follow a similar trajectory.
Sequoia frames AI transformation as "service-as-a-software." The market isn't just competing for software budgets. It's targeting the multi-trillion-dollar services market. Sequoia contrasts this with cloud-era seat-based pricing, arguing that AI companies will increasingly "sell work ($/outcome)" rather than "sell software ($/seat)." Sierra, for example, charges per resolution.
a16z points out that the key shift in enterprise software is turning passive database systems into something closer to autonomous workflow engines.
The Numbers
- Mordor Intelligence estimates the Workflow Automation market at $26.01B in 2026, growing to $40.77B by 2031 (CAGR 9.41%)
- Precedence Research projects the global enterprise software market (traditional seat-based/subscription software) to reach $761.73B by 2034
- Precedence Research projects the global BPO (business process outsourcing) market to reach $840.60B by 2034, up from $315.46B in 2024
My Lens
The numbers show we're still in the early stages, similar to the iPaaS phase. Fortune Business Insights defines iPaaS as a cloud service that enables platforms and applications to integrate and exchange data (essentially "integration + automation pipelines").
From my observation, this wave of rapid revenue growth caught enterprises that hadn't digitally transformed during COVID. AI workflows gave them a reason to finally restructure their entire operational layer.
AI-native enterprise starts with workflow automation but integrates three major markets (which is why I chose these three data points). If you dig deeper, the logic is: use "data systems" as the foundation, "execution systems" as the product, and "outcomes" as the pricing unit.
Databases as the storage layer rarely undergo significant migration. This means legacy providers like Oracle and Salesforce still have strong moats. Meanwhile, outcome-based pricing models are essentially seeking new engagement patterns in the BPO market, like Sierra's per-resolution billing.
The contract-based market will see a new wave of design, including human-in-the-loop copilots and more autonomous autopilots. Different industries (finance, healthcare, etc.) will dictate different product choices.
Who will be the next standout unicorn after Sierra?
Creative Generation
YC's RFS highlights "Video Generation as a Primitive" that will reshape media/entertainment, shopping, and gaming/simulation.
a16z repeatedly positions video as the creative and media format of choice (including "creative tools" and "step inside video"), and notes how multimodal capabilities enable new uses for "video data."
The Numbers
- Dimension Market Research estimates Generative AI in media & entertainment growing from $1.9B in 2024 to $16.8B by 2033
- Digital Content Creation market projected to reach $90.4B by 2033 (CAGR 12.8%)
- Mordor forecasts the video game market at $289.73B in 2025, growing to $531.77B by 2030
My Lens
The market looks compelling, both in numbers and potential. It feels like the start of a new era. But whether we're talking about media, shopping, or gaming, user acceptance still has a long way to go.
The technology can already deliver usable results that give users aha moments ("this is generated by AI?"). But take AI actors, for example. Audience sentiment (not from actors, but from viewers) shows emotional resistance, even though many can barely distinguish whether media (images, videos) is AI-generated anymore.
Then there's gaming. I read an article exploring why even the simplest NPCs haven't adopted AI for dialogue yet. Commenters joked about LLMs ("Yes, you're absolutely right," though the game script doesn't allow that direction). The deeper issue isn't just about implementation. It's about redesigning game structure, rethinking interaction models, and flipping game logic.
This explains why simulation has seen more traction in industrial applications than in entertainment.
Reading the Signals
Across these three reports, a clear pattern emerges around timing. The momentum right now is in infrastructure over applications. Multi-agent orchestration, workflow engines, and the foundational systems are getting built and shipped fast.
Enterprise software still has significant room to develop. Most companies are just beginning to figure out what AI-native operations actually look like.
Creative generation shows promise in the numbers, but multimodal capabilities and world models need more time. The technology works, but user acceptance lags, especially in entertainment and media.
When application startups keep getting wiped out by big company product launches, is infrastructure where the real opportunity lies?