How AI Agents Work, and How to Secure Them
An agent is a model that directs its own tool use. That flexibility is useful and risky. Here is how agents differ from workflows, when to use one, and the controls that matter.
Explainer4 min read
What matters today
- 01How AI Agents Work, and How to Secure Them
An agent is a model that directs its own tool use. That flexibility is useful and risky. Here is how agents differ from workflows, when to use one, and the controls that matter.
- 02What Is Model Context Protocol (MCP)? A Practical Guide
MCP is an open standard for connecting AI applications to tools and data. Here is what it defines, what it leaves to you, and where it can go wrong.
- 03Inference Serving Explained: Batching, KV Cache, Quantisation and Speculative Decoding
Serving a language model is mostly a memory and scheduling problem. Four techniques explain most of why one GPU can serve many users, and what each one costs.
- 04How to Check an AI Startup Funding Announcement Against SEC Form D
A funding headline is a company's own statement. For US private raises, the SEC's Form D is a second public record you can compare it with.
Trending now
Where the stories are
Latest
Fresh from the desk
Startups
How to Check an AI Startup Funding Announcement Against SEC Form D
A funding headline is a company's own statement. For US private raises, the SEC's Form D is a second public record you can compare it with.
Guide3 min read
Models
Meta Muse Wonder: Specs, Pricing and How to Read Its Benchmark Claims
Meta's first model it calls frontier-class is available by API and by subscription. The documentation is clear on price and context; the model card is more revealing about how far its benchmark comparisons can be trusted.
Analysis7 min read
Robotics
Vision-Language-Action Models: How Robots Borrow From Language Models
A vision-language-action model takes a camera image and an instruction and outputs robot actions. Three papers show how that idea was built, opened up and extended.
Explainer4 min read
Agents
AI Agent Memory: Notes, Compaction and Context Editing Compared
An agent that runs for hours outgrows its context window. Three mechanisms keep it going, and each one forgets differently.
Explainer6 min read
Agents
Multi-Agent Systems: When Subagents Help and When They Hurt
Two well-known engineering posts from June 2025 reach opposite conclusions about multiple agents working together. Both are right about their own task. Here is the decision rule that reconciles them.
Guide6 min read
Topic radar
Follow the threads
- 16 storiesModelsFoundation models, launches, capabilities, pricing and how they compare in practice.
- 5 storiesAgentsAutonomous and semi-autonomous systems: tool use, orchestration, memory and evaluation.
- 6 storiesResearchPapers and results that change what is possible, read closely and explained plainly.
- 17 storiesDeveloper ToolsSDKs, coding assistants, evaluation harnesses and the tooling AI engineers rely on.
- 2 storiesEnterprise AIAdoption, procurement, ROI and the operational reality of AI inside large organizations.
- 1 storyOpen SourceOpen-weight models, repositories, licenses and the community building in public.
- 2 storiesSafetyAlignment, evaluations, security and the work of making AI systems reliable.
- 3 storiesPolicyRegulation, standards and government action that shape how AI is built and used.
- 1 storyRoboticsEmbodied AI, manipulation, autonomy and the hardware that connects models to the world.
- 2 storiesInfrastructureCompute, chips, inference serving, data centers and the economics of running AI.
- 1 storyStartupsNew companies, funding and products worth watching, with the claims checked.
Deep reads
Worth the long read
Models9 min read
Claude Haiku 5.5: Specs, Pricing, Migration Traps and When to Use It
Anthropic's smallest current model costs about a twentieth of Sonnet 5.5 per token, but it has a price cliff at 100,000 tokens and five breaking API changes. What the documentation says, and what it means for routing.
- 02
Reasoning Models Explained: Thinking Tokens, Effort Settings and What They Cost
A reasoning model spends tokens thinking before it answers. What that buys, how Anthropic and OpenAI expose and bill it, how effort settings work, and why you should not treat the visible reasoning as proof.
Models · · 8 min read
- 03
How to Build LLM Evals That Survive Model Changes
Public benchmarks tell you how a model does on someone else's tasks. An eval suite built from your own tasks tells you whether a new model, prompt or price tier is safe to ship.
Research · · 8 min read
Open source and developer pulse
What builders are shipping
- InfrastructureInference Serving Explained: Batching, KV Cache, Quantisation and Speculative DecodingSource: arxiv.org ·
- ModelsMeta Muse Wonder: Specs, Pricing and How to Read Its Benchmark ClaimsSource: research.meta.ai ·
- AgentsAI Agent Memory: Notes, Compaction and Context Editing ComparedSource: platform.claude.com ·
- AgentsMulti-Agent Systems: When Subagents Help and When They HurtSource: anthropic.com ·
- ModelsReasoning Models Explained: Effort, Thinking Tokens and What They CostSource: platform.claude.com ·
Each line links to our coverage. The source shown is the first one cited in that story. We do not display repository statistics we have not verified.
The ANTM newsletter
The signal, not the noise.
- A short daily or weekly briefing: what happened, why it matters, who is affected.
- Every claim linked to its primary source and checked before it ships.
- No hype, no filler. Unsubscribe in one click.