88,6%Claude Opus 4.8 scores 88.6% on SWE-bench Verified, versus about 4% for 2023 models

71%developers who regularly use AI agents work in Claude Code

1M tokensOpus 4.8, Sonnet 4.6, and Fable 5 have the context - the agent can keep a large repository in memory

30+agent platforms execute the same skill in SKILL.md format without rewriting

Industry solutions

What Can Be Done with AI-native Development

Corporate website and marketing A steady flow of pages, case studies, landing pages, and bilingual versions The content manager edits Markdown, the skill builds the page with SEO and clean URLs, and the EN version is generated automatically

E-commerce and marketplaces Showcases, product cards, catalog synchronization across systems Skills for standard pages plus MCP for PIM/OMS/ERP instead of manual integrations for every change

Manufacturing Technical catalogs, documentation, specifications Skills encapsulate data formatting rules; the agent keeps a large catalog in a 1M-token context

Distribution and B2B B2B portals, supplier onboarding, shared catalogs Reusable components and skills for portals; the expertise is transferable and handed over to the client

Finance and Insurance Product pages, calculations, and documents with strict data requirements MCP with source access control; quality and compliance checks are built into the build

Media and content projects High-frequency publishing of content and special projects Skills for publishing formats plus automatic SEO and visual audits on every build

Internal portals and documentation Operating procedures, knowledge bases, living technical documentation Living knowledge base: the agent keeps schemas and contracts from the repository up to date

Capabilities

AI-native development capabilities

Prompt: business task (new page, case study, integration, fix)The agent selects the skill that encapsulates the solution algorithmReading context from the repository: narrative, components, procedures, historyAccess to data and systems through MCP - without custom connectorsAssembly from reusable components following the skill stepsAutomated build checks: SEO, visuals, copy, clean URLsPublish and live-check; the expertise remains in the skills for the next task

The task is defined by a prompt; the agent selects the needed skill, reads context from the repository, accesses systems through MCP, assembles the result from reusable components, passes automated checks, and is published. The expertise stays in the skills and repository and is reused on the next task.

Speed to Result

A standard task (new page, case study, block, SEO check) is solved in a few prompts because the algorithm is already built into the skill instead of being reconstructed each time

Reproducibility

The skill defines the same execution path for any operator, so the result does not depend on who is using the agent

Corporate Memory

Narrative, guidelines, the design system, and solved tasks are stored in the repository and skills, not in people’s heads - knowledge accumulates instead of leaking when people leave

Transferability

The repository, skills, and MCP servers are handed over to the client in full; another team can continue the work without archaeology

No vendor lock-in

Skills (SKILL.md) and integrations (MCP) are open standards; you can change the model and vendor without rewriting the expertise

Quality as part of the process

Validators, audits, and critics for copy, visuals, and SEO are built into the repository and run on every build, instead of relying on human review

Lower Barrier to Entry

The content manager edits Markdown and triggers publishing independently - without handing every change to development

Bilingual by default

RU and EN versions are generated from a single source with explicit overrides - no second team and no second content tree

Approach

How we implement AI-native development

Without modifying the core

We do not fork or patch the core of AI-native development. AI-native development stays on the standard updatable version - we move business logic into separate microservices alongside it, so platform updates do not break your custom work.

International Standards, Not Homegrown Hacks

Where a mature international solution exists, we use it instead of inventing our own protocol or platform. Before writing code, we study how the problem is already solved in the industry.

Transferability

The solution is loosely coupled and documented: it can be handed over between teams and contractors without rewriting. You are not tied to us.

AI compatibility

AI-native development within the AI stack

Agent

Claude Code / Codex / Cursor with access to the repository and tools; context up to 1M tokens

Repository

Single source of truth: content, components, narrative, guidelines, and history - all versioned

Skills (SKILL.md)

A folder with Markdown instructions and scripts; they encapsulate expertise, load on demand, and follow an open standard

MCP

An open protocol for connecting data and systems: one interface instead of N×M connectors, governed by the Linux Foundation

Engineering foundation

Loose coupling and transferability, designed for agents: the agent speeds up work on healthy code

Build Checks

Validators and SEO/visual/copy audits in the repository block publishing poor-quality output

Context 2026

What changed in the market

Agentic coding has become the norm

By 2026, AI agents (Claude Code, Codex, Cursor) understand repositories, edit files, run tests, and iterate; Claude Code leads, with 71% of AI-agent users relying on it (Pragmatic Engineer Survey, 2026)

Engineering quality of models

Claude Opus 4.8 (May 28, 2026) scores 88.6% on SWE-bench Verified, compared with about 4% for 2023 models; on June 9, 2026, Anthropic released Claude Fable 5, its most powerful public model ($10/$50 per 1M tokens)

Context: 1M tokens

Opus 4.8, Sonnet 4.6, and Fable 5 work with 1 million tokens of context at standard pricing, with no long-context premium, allowing the agent to keep a large repository in memory in full

Skills: an open standard

On December 18, 2025, Anthropic released the Agent Skills specification (SKILL.md, a folder with Markdown files and scripts); Microsoft, OpenAI, Cursor, GitHub, and Figma adopted it, and the skill works across 30+ platforms

MCP: neutral connectivity

Model Context Protocol, launched in November 2024, was supported by OpenAI, Google, and Microsoft by 2025; in December 2025, Anthropic transferred the protocol to the Linux Foundation (Agentic AI Foundation), reinforcing its neutrality

Honestly

Pros and cons

Pros

Complex routine tasks are solved in a few prompts: the expertise is captured in skills instead of being repeated manually each time
The result is reproducible: the skill defines a single algorithm regardless of who executes it
Company knowledge accumulates in the repository and skills instead of being lost when people rotate out
Everything is transferable: the repository, skills, and MCP integrations are handed over to the client and can be continued by another team
No vendor lock-in - SKILL.md and MCP are open standards, and the model or vendor can change without rewriting expertise
Quality is checked automatically during build instead of depending on the reviewer’s attention
The barrier to entry drops: a content manager can publish edits independently, without escalating to development

Cons

Without an engineering foundation, an agent speeds up degradation: AI code tends to duplicate itself (GitClear: repeats x8, copy-paste 8.3% to 12.3%, refactoring below 10%)
Hidden technical debt: working but inconsistent pieces accumulate quietly and later shift the team's time from building to fixing
Upfront investment is required: architecture, the skill library, and operating procedures must be designed before the agent starts working
The agent needs verification: on complex and atypical tasks it can make mistakes, so validators and human review of key decisions are mandatory
The cost of runs on high-end models and large contexts is significant, so it must be managed deliberately (effort, task budgets)
The benefit shows up on a stream of repetitive tasks; for a one-off unique job, the overhead of skills may not pay off

Blog