← /editions
#13

The model isn't the bottleneck

// Three builders shipped agent infrastructure rather than models — a free 200-page agentic security reference, a case for governed agentic data pipelines, and an 'OS' for deploying agents into chat — converging on one lesson: capability is cheap now, while trust, controls, and grounding are scarce.

$ today_in_10_seconds

Three builders shipped infrastructure today, not models — and together they make one point: the hard part of agentic AI is the scaffolding around the agent, not the agent itself. A free 200-page security reference, a case for governed agentic data pipelines, and an 'OS' for deploying agents into chat channels all converge on the same lesson: capability is cheap now; trust, controls, and grounding are what's scarce.

$ lead_story

A free, 200-page agentic AI security reference (CC BY-NC-ND 4.0)

Fernando Lucktemberg released The Agentic AI Security Stack, a 200+ page practitioner reference that's free, needs no email, and carries a Creative Commons license so teams can share it without a legal review step. What sets it apart from typical 'sanitize your inputs' guidance is method: it starts from a threat model, traces a single attack kill-chain through twelve interception points — threat modeling, credential architecture, isolation, egress control, human-in-the-loop gates, memory integrity, orchestration trust, agent-native identity, MCP protocol hardening, supply-chain verification, audit logging, and regulatory compliance — and maps every control to OWASP, MITRE ATLAS, and CSA MAESTRO. One sharp distinction it draws: OAuth with PKCE verifies you're talking to the authorized MCP server, but does nothing against tool-poisoning — a legitimate server updated maliciously passes OAuth while delivering injected instructions. For anyone connecting agents to external tools, that gap is the ballgame.

// why_it_matters: Agentic systems are shipping with structural vulnerabilities most teams haven't named yet — and this is the first free reference that starts from the attacker's kill-chain instead of a control checklist.

read more →

$ also_worth_attention

  • Agentic data engineering: the model isn't the bottleneck, the harness is

    RevOS's Pavel Nedelko makes the sharpest version of an argument every agent builder needs to internalize: an AI that can write data code and one you can trust to ship it are different things, and the gap is the harness — grounding plus controls. The proof point is concrete: a general model scored ~51% on Snowflake's text-to-SQL evaluation, but grounding the same task in a governed semantic layer pushed accuracy past 90%. The operating principle to steal: the agent fails at review time, not in production — changes ship as pull requests that pass tests and CI before a human merges.

    link →
  • "Agentic OS" — deploying agents into the channels teams already use

    Injecting.ai's Agentic OS pitches itself as a deployment layer that drops AI agents into Telegram, Slack, and Discord with tools, knowledge, and code execution on one platform — meeting users where they already work rather than in a bespoke UI. It lands in a crowded race to define the 'agentic operating system,' from Salesforce recasting Slack as one to open frameworks like Agno. The unresolved tension worth watching: whether an 'OS' that lives inside someone else's chat product is infrastructure you own, or a feature set you rent.

    link →
  • The "OS for agents" land-grab is on

    Zoom out and today's third story is part of a pattern: everyone from Salesforce to Microsoft to indie builders is claiming operating-system status for their agent layer. Critics argue the label demands real properties — multi-model support, system-level integration across your stack, and data sovereignty — not just AI features bolted onto existing SaaS. The strategic question for buyers isn't which vendor's agent features are best; it's whether to rent agent capabilities inside a walled garden or own the orchestration layer outright.

    link →

$ bigger_picture

The trend behind the headlines

Notice what nobody shipped today: a model. All three stories are about the layer around the agent — security controls, data grounding, deployment surface — and that's the tell. As frontier and open-weight models converge on similar capability, the durable engineering problems are migrating outward: how do you stop a poisoned tool from hijacking an agent, how do you ground it so its output is right and not just fluent, and where does it actually run. The recurring primitive across all three is human-in-the-loop review as a hard technical gate — the security book makes it a chapter, the data-engineering piece makes it the whole trust model ('fails at review time'), and the OS debate is really about who controls that gate. For builders, the takeaway is that your competitive edge is increasingly the harness, not the model you call.

$ tool_of_the_day

MCP (Model Context Protocol) — and hardening it

MCP shows up in all three of today's stories as the connective tissue between agents and tools, which is exactly why it's becoming the highest-value attack surface. If you're wiring agents to external servers, pair the protocol with the security reference's Chapter 9 guidance: treat transport security (is this the right server?) and content security (is the tool description free of injected instructions?) as separate problems, because OAuth solves only the first. Start at the spec, then harden deliberately.

explore MCP (Model Context Protocol) — and hardening it →

$ subscribe --free

Get the next edition in your inbox

// DAIly Agent covers agentic AI three times a week

// free, no spam, exit anytime