The Landscape

Why it matters

Claude is quietly becoming the default model for serious agent builders. The combination of massive context windows, strong tool use, and reliable instruction following makes it the backbone of most production agent systems i see shipping today. Claude Code is also pushing the boundary on what a coding agent can actually do in practice.

Why it matters

Still the default for most developers. The Assistants API and function calling are how most people build their first agent. But the gap with competitors is narrowing fast, and i'm seeing more teams diversify away from OpenAI-only stacks. The o-series reasoning models are genuinely impressive for complex multi-step tasks though.

Why it matters

Google has the distribution advantage nobody else can match. A billion people already use their products. If Gemini gets good enough (and it's getting there), the default agent for most consumers will just be.. Google. The Gemini CLI launch was a smart move, giving developers a free on-ramp.

Why it matters

Llama changed the game for anyone building agents who can't or won't depend on API providers. The open-source ecosystem around it is massive. But here's the honest take: for most production agent use cases, the hosted APIs (Claude, GPT) still outperform on reliability. Llama is great for cost-sensitive deployments and edge cases where you need full control.

Why it matters

DeepSeek is the most important thing that happened to AI economics in 2025. They proved intelligence is deflationary. This has massive implications for agent builders because cheaper inference means agents can reason longer, run more checks, handle messier tasks. The cost floor keeps dropping.

Why it matters

Mistral matters for the agent ecosystem because they're proving that European AI can compete on model quality. For teams building agents that need to comply with EU regulations, having a strong European model provider is genuinely useful. Their function calling support is solid.

Why it matters

The interesting angle here isn't the model quality (it's competitive but not leading). It's the data advantage. Having real-time access to X's firehose of social data gives Grok-powered agents a unique edge for anything involving current events, sentiment, or social signals.

Why it matters

Cohere is playing a different game than the consumer-facing model providers. They're going deep on enterprise, and that focus shows in their RAG capabilities and data privacy features. If you're building agents for regulated industries, they're worth a look.

Why it matters

Love it or hate it, LangChain is where most people start. The ecosystem is massive, the docs are good, and LangGraph genuinely solves hard orchestration problems. The learning curve is steeper than alternatives but it handles complex workflows better than anything else i've used. If you need state machines and branching logic, start here.

Why it matters

CrewAI is the fastest path from idea to working multi-agent prototype. You can get something running in a day. The role-based metaphor (agents as team members with specializations) is intuitive and maps well to how people think about delegation. It starts to creak under production load, but for learning and prototyping, nothing's faster.

Why it matters

AutoGen's conversation-based approach to multi-agent systems is genuinely different from the alternatives. Agents talk to each other to solve problems, which creates interesting emergent behaviors. The Microsoft backing means solid enterprise integration, but the developer experience still lags behind LangGraph and CrewAI.

Why it matters

This is OpenAI saying the framework layer matters enough to own it themselves. The Agents SDK is clean and well-designed, but the obvious limitation is vendor lock-in. You're building on OpenAI's stack, with OpenAI's models. For teams already committed to the OpenAI ecosystem, it's the smoothest path. For everyone else, it's a question of how much lock-in you're comfortable with.

Why it matters

If you're a Python developer who cares about type safety and clean code, this is probably the framework that'll feel most natural. The Pydantic team knows how to build developer tools. It's newer and less battle-tested than LangChain but the DX is noticeably better for structured agent outputs.

Why it matters

LlamaIndex carved out a clear niche: if your agent needs to work with your data, this is where to start. The RAG pipeline tooling is best-in-class. It's less of a general-purpose agent framework and more of a data layer that agents can plug into. The two aren't mutually exclusive, many teams use LlamaIndex for data + LangGraph for orchestration.

Why it matters

If you're a .NET shop, this is your only real option for agent building, and it's actually quite good. The enterprise focus means solid patterns for things like authentication, logging, and observability that other frameworks treat as afterthoughts. Not the sexiest framework, but production-ready.

Why it matters

Manus felt like the ChatGPT moment for agents. The frenzy was real: invite codes selling for thousands of dollars, Discord swelling past 180K members in days. i did a deep dive on them. The honest assessment: the underlying architecture (Claude + Qwen running in sandboxed VMs) isn't revolutionary, but the product execution is. The question is whether reliability can scale. Early users reported looping errors and lazy behavior on longer tasks. But $90M ARR in 5 months doesn't lie.

Why it matters

Perplexity is the clearest example of an agent that regular people actually use. The deep research feature is genuinely impressive for complex questions. Give it a real research problem and compare to an hour of Googling. The quality gap is significant. This is what consumer-facing agents should feel like.

Why it matters

ElevenLabs is making voice the command layer for agents, and the math is compelling: we speak at 150 words per minute, type at 40. 11ai turns research, execution, and multi-step tasks conversational. With MCP integration, it plugs into tools people already use. i covered their launch in Agent Angle #1. Voice-first agents feel like a genuinely different interaction paradigm.

Why it matters

HeyGen is pushing video from a creative tool into an agent workflow. Instead of hiring actors and studios, you describe what you want and an agent generates it. The quality is getting eerily good. The interesting question is what happens when video agents get as good as text agents, because most of the internet is still consumed as video.

Why it matters

i did a deep dive on Decart. An investor told them it was impossible, but they built a system that generates interactive worlds 10x cheaper than existing methods. The implications for gaming, simulation, and training environments are massive. If agents need to practice in simulated worlds before acting in real ones (and they will), Decart's approach could be foundational.

Why it matters

Cursor is the coding agent most developers actually use daily. It's not the most autonomous (Claude Code goes further) but the integration into the editing workflow is seamless. The agent mode is genuinely useful for medium-complexity refactoring tasks. i think the IDE-native approach is the right bet for most developers.

Why it matters

Devin's launch in early 2024 was a watershed moment. It showed people what a fully autonomous coding agent could look like. The reality has been more nuanced. Production reliability is still a challenge, and most teams find they get better results from copilot-style tools (Cursor, Claude Code) than fully autonomous agents. But Devin pushed the whole space forward.

Why it matters

i've been using Claude Code heavily and it's shifted how i think about what coding agents can do. The terminal-first approach means it has real access to your tools and environment, not just a sandboxed editor. The human-in-the-loop design (you approve each action) trades autonomy for reliability, and for production code, that's the right tradeoff.

Why it matters

Replit Agent is interesting because it controls the full stack, from code to deployment. That closed-loop advantage (sound familiar?) means it can actually ship working apps, not just write code. The tradeoff is you're locked into Replit's platform. i covered a case where it deleted live data, which highlights the reliability gap that all coding agents face.

Why it matters

Copilot has the distribution advantage. Tens of millions of developers already use GitHub. The agent mode is catching up to Cursor but the real moat is the GitHub integration: issues, PRs, code review, Actions. If Copilot agents can operate across the full GitHub workflow, that's hard to compete with.

Why it matters

The Windsurf acquisition by OpenAI tells you everything about how important the coding agent space has become. OpenAI wanted an IDE and they bought one. For users, this means Windsurf will likely get deep GPT integration that other editors can't match. The competitive dynamics between Cursor/Windsurf/Copilot are worth watching closely.

Why it matters

v0 is interesting because it's targeting a very specific workflow (frontend generation) and doing it well. The code quality is surprisingly good for a generative tool. It's a sign that the best agent products will be narrow and excellent rather than broad and mediocre.

Why it matters

bolt.new and Lovable represent the zero-to-deployed agent experience. No terminal, no IDE, just describe what you want. The target audience isn't developers; it's everyone else. The quality ceiling is lower than Cursor/Claude Code for complex apps, but for getting something live fast, the friction reduction is real.

Why it matters

Same thesis as bolt.new: the next wave of builders won't write code. They'll describe what they want and an agent builds it. Lovable is one of the better executions of this idea, with solid design defaults and Supabase integration for the backend.

Why it matters

i did a full deep dive on Sierra. They're building an outcome-priced agent OS for customer service, betting that reliability engineering and supervision can make probabilistic systems dependable enough for Fortune 500 scale. The founding team is absurdly credentialed. The key insight is they're not replacing customer service reps; they're replacing the vending machine experience of bad support. Early results from their enterprise deployments are genuinely impressive.

Why it matters

Fin has a distribution advantage that pure-play agent companies don't: Intercom already sits in the customer support workflow for thousands of companies. Adding an agent layer on top of existing conversations, knowledge bases, and routing is a natural extension. The resolution rates are competitive with specialized agent companies.

Why it matters

Ada has been in the automated support game longer than most, and the shift from scripted chatbots to reasoning agents has been good to them. They already had the integrations, the customer relationships, and the conversation data. Adding LLM-powered reasoning on top of that existing foundation is exactly the closed loop advantage i keep talking about.

Why it matters

Same thesis as Salesforce: the companies that already own the workflow have the best shot at adding agents on top. ServiceNow's IT service management agents are one of the most proven enterprise agent use cases. When an employee reports a laptop issue, an agent that can check inventory, create a ticket, and schedule a replacement is genuinely useful.

Why it matters

Legal is one of the strongest verticals for AI agents because the work is high-value, document-heavy, and follows patterns. Harvey's moat is the legal training data and the relationships with major firms. A coding agent can be swapped easily. A legal agent that understands specific jurisdictional requirements and has been validated against real case law? That's sticky.

Why it matters

Glean is solving the internal knowledge base Q&A problem, which is one of the most proven agent use cases for enterprises today. If you have good source material and fragmented knowledge across tools, an agent that can search everything and synthesize answers is genuinely useful. Their connectors to enterprise apps are the moat.

Why it matters

Hebbia is going after the analyst workflow in finance and legal, where people spend hours reading documents. Their Matrix product lets you run structured analysis across thousands of documents simultaneously. The bet is that the best AI companies will be the ones that own specific high-value workflows, not general-purpose chat.

Why it matters

Klarna is the poster child for the agent economics argument. They cut their workforce and claimed the AI handles conversations better, faster, and cheaper. Whether you find that inspiring or unsettling, the numbers are hard to argue with. It's also a cautionary tale: Klarna owns the full stack (customer data, transaction history, support workflows), which is why it works. Most companies trying to replicate this don't have that advantage.

Why it matters

i did a deep dive on TinyFish. Most of the web is a mess, tables that don't render, dynamic content, pages built for humans not machines. TinyFish uses AI to read the unreadable web and turn it into the world's most valuable database. Data extraction infrastructure like this is the picks-and-shovels play for the agent economy. Every agent that needs to understand the web needs something like this.

Why it matters

Nvidia is winning right now, but 90% gross margins never last forever. Major customers (Amazon, Google, etc.) are building their own chips. At the same time, more efficient models like DeepSeek mean you need fewer GPUs to get the same work done. This challenges the infinite demand thesis. But for now, if you're running agents at scale, you're running on Nvidia hardware.

Why it matters

Every agent that needs to interact with the web needs a browser. Browserbase is the infrastructure layer that makes this work at scale. Reliable browser automation is one of those unsexy problems that turns out to be critical for production agents. The picks-and-shovels play for the agent economy.

Why it matters

The web wasn't built for AI agents to read. Firecrawl translates messy web pages into clean data that agents can actually work with. It's a simple tool that solves a real problem. If you're building any agent that needs to pull information from the web, you'll probably end up using something like this.

Why it matters

Regular search APIs return results designed for humans to browse. Tavily returns results designed for agents to consume. It's a small but important distinction that matters a lot when you're building agents that need to research and gather information. The fact that it's become the default search tool in LangChain says something.

Why it matters

If you don't log everything your agent does, you're flying blind. i keep saying this: log every tool call, every model response, every decision point. LangSmith makes this practical. The evaluation features for comparing agent performance over time are also genuinely useful for closing the feedback loop.

Why it matters

Braintrust tackles the evaluation problem head-on. Most agent benchmarks are broken (i wrote about this). What you need is evaluation tied to your specific use case, with your actual data, measuring metrics that matter to your users. Braintrust provides the infrastructure for that kind of real-world evaluation.

Why it matters

W&B has been the standard for ML experiment tracking for years, and they've adapted well to the LLM era. The Weave product for LLM application monitoring is solid. For teams that are already using W&B for model training, adding agent monitoring is a natural extension.

Why it matters

Helicone's value proposition is simplicity. One line of code to start logging all your LLM calls. For teams that don't need the full weight of LangSmith or W&B, Helicone is the fastest path to visibility into what your agents are actually doing and costing.

Why it matters

Apify has been doing web scraping longer than the AI agent wave, and that experience matters. Their marketplace of pre-built scrapers means agents can plug into thousands of data sources without building custom integrations. For data-hungry agents, Apify is often the fastest path to getting the data they need.

Why it matters

When agents write and execute code, you need sandboxing. E2B provides the secure execution environments that let agents run code without nuking your production database. After the stories i've covered of agents deleting live data and firing root commands, sandbox infrastructure feels less like nice-to-have and more like essential.

Why it matters

Every agent needs tools, and building tool integrations from scratch is tedious, repetitive work. Composio pre-builds the integrations so you can focus on the agent logic. For prototyping especially, being able to plug in 250+ tools immediately is a huge time saver.

Why it matters

Scale AI is the invisible infrastructure behind many of the models powering agents. Their data labeling and RLHF work feeds directly into model quality. The government contracts are also interesting: if agents are going to operate in government contexts, the evaluation and safety infrastructure Scale provides becomes critical.

Why it matters

Healthcare and biotech are the verticals where AI agents could have the most impact. Chai-2 is rewriting the biotech R&D playbook by letting agents do the kind of molecular analysis that used to take teams of PhDs months. The compliance requirements in healthcare create a natural moat. Regulated industries are where vertical agents win.

Why it matters

Salesforce has the enterprise distribution that startups can only dream of. Agentforce agents sit inside existing Salesforce workflows with access to customer data that's been accumulating for years. That's a massive closed-loop advantage. The question is whether Salesforce can execute on the AI product fast enough, or if nimble startups eat their lunch.

Why it matters

Tesla's advantage in robotics is the same as their advantage in EVs: vertical integration and manufacturing scale. If anyone can bring the cost of humanoid robots down to mass-market levels, it's Tesla. The factory deployments are early but real. i covered K-Scale's attempt to compete with Tesla on open-source robots.. the economics are brutal. Tesla's scale is nearly impossible to match.

Why it matters

Figure has raised more money than almost any robotics startup in history, and the OpenAI partnership gives them access to frontier AI capabilities. The Figure 02 demos showing natural conversation while performing physical tasks are impressive. But robotics is hard. Really hard. The gap between demo and production is even wider than in software agents.

Why it matters

i wrote a deep dive on K-Scale. They built an open humanoid faster than Tesla, which is genuinely remarkable. But their story also reveals the fragile economics of open-source robots. Hardware margins are thin, iteration is slow, and the capital requirements are enormous. What their failure tells us is that open-source might work for software agents but the economics don't translate to hardware.

Why it matters

Unitree is doing to robots what Chinese manufacturers did to EVs and drones: making them dramatically cheaper. A $16K humanoid robot changes the math on what deployments are economically viable. The quality question remains, but if Unitree can deliver reliable robots at these price points, it changes the entire competitive landscape.

Why it matters

1X is taking a different approach to humanoid robots: safety-first design for human environments. While Tesla and Figure target industrial settings first, 1X is designing for homes and offices from the start. The safety-first approach could be a genuine differentiator as robots move beyond controlled factory floors.

Why it matters

Agility is one of the few robotics companies that can point to real commercial deployments, not just demos. The Amazon partnership for warehouse operations is a meaningful validation. Digit isn't trying to do everything; it's focused on the specific physical tasks that matter in logistics. That vertical focus is the same pattern i see winning in software agents.

Why it matters

Boston Dynamics proved that extreme physical capability is possible, but struggled for years to turn it into a business. The Hyundai acquisition and commercial pivot are changing that. Spot has real commercial deployments in inspection and surveying. The lesson here: in robotics, the gap between can we build it and can we sell it profitably is enormous.

Why it matters

Physical Intelligence is trying to do for robot control what foundation models did for language: build one model that generalizes across many tasks and hardware platforms. If they succeed, it changes the economics of robotics completely. Instead of programming each robot for each task, you fine-tune a foundation model. Early results are promising but it's still very much research-stage.

Why it matters

Apptronik's modular approach is smart. Instead of one robot for everything, they build a platform that can be configured for specific tasks. The NASA and Mercedes partnerships validate the hardware quality. Whether that's enough to compete with Tesla's scale and Unitree's pricing remains the question.

Why it matters

Sanctuary AI's emphasis on hand dexterity is key. As i wrote in the Dexterity Stack deep dive, intelligence lives at the point of contact. A robot that can fold towels, handle irregular objects, and adapt to unstructured environments solves the hardest problems in physical AI. Sanctuary is one of the few teams seriously tackling this.

Why it matters

Virtuals is the most interesting experiment in on-chain agents right now. The idea of tokenized, community-owned agents is genuinely novel. Whether it's a durable business model or a speculative vehicle remains to be seen. The token economics create incentives for agent development that don't exist in traditional software. But the space is still.. very early and very wild.

Why it matters

ELIZA's journey from meme to legitimate framework is one of the stranger stories in the agent space. The open-source community around it is genuinely active, and the agent capabilities (social posting, trading, cross-platform interaction) are real. But the crypto-native context means everything is wrapped in speculation and token dynamics that make it hard to evaluate on pure fundamentals.

Why it matters

Fetch.ai has been in the AI-meets-crypto space longer than most. The merger with SingularityNET and Ocean Protocol creates the largest decentralized AI ecosystem. The thesis is that autonomous agents need decentralized infrastructure for identity, payments, and coordination. Whether that thesis is correct is still an open question.