.jpg)
Here is a number worth sitting with: 42% of companies abandoned most of their AI initiatives in 2025 — up from just 17% the year before — with an average sunk cost of $7.2 million per failed project. And yet, in the same year, AI investment budgets kept growing. The problem is not enthusiasm. The problem is architecture.
The companies that deploy AI agents and see them fail in production are not using bad models. In most cases, the model performs exactly as advertised. What breaks is everything underneath it — the structural layers that determine whether an agent understands your business, retains context, executes safely, and scales without turning into a liability.
This article breaks down the seven architecture layers that production-ready AI agents require. If you are evaluating whether to build an agentic system, or trying to understand why the one you shipped is underperforming, this is the framework your team needs before writing another line of code.
The industry spent years framing agentic AI as a model selection challenge. Pick the right LLM, engineer a strong prompt, and the agent would figure out the rest. That assumption has now been stress-tested at scale — and it has not held up.
According to Gartner's 2026 analysis, only 23% of failed AI initiatives trace their failure back to model underperformance. The other 77% trace to strategy, governance, and architecture that was never designed to support autonomous systems. A separate Gartner projection warns that 40% of enterprise AI agent projects will be cancelled by 2027 — not because the technology is immature, but because the underlying architecture was not built to support it.
A useful reframe comes from researchers at Halmstad University, who concluded that reliability in agentic systems is "first and foremost, an architectural property." It is not something that emerges from a better model. It emerges from how a system is deliberately decomposed into components, and whether each component is designed to handle real-world load, failure, and ambiguity.
The distinction between an AI chatbot and an AI agent makes this concrete. A chatbot responds to input and completes a single turn. An AI agent plans, acts across multiple steps, loops back on its outputs, and self-corrects when something goes wrong. That is a fundamentally different system — one that needs a fundamentally different infrastructure.
The following framework maps out the structural layers that every production AI agent requires. Every layer is load-bearing. Miss one and you create a system that either fails outright or becomes unpredictable under pressure.
The first layer is the interface between the agent and the humans it serves. This goes beyond the UI. It includes the feedback loops that allow users to set goals, review outputs, and correct the agent's direction over time.
Most teams skip this layer entirely at the design stage. They build the agent's capabilities before defining how users will actually interact with it, what kind of feedback the system will capture, and how that feedback loops back into agent behavior. The result is an agent that operates blind to what the business actually needs — technically functional, strategically useless.
Good user interaction design for agentic systems includes clearly defined input mechanisms, structured output review flows, and explicit checkpoints where a human can redirect or override the agent before it proceeds. Skipping this is not a shortcut. It is a guaranteed path to an agent that diverges from business intent within weeks of deployment.
An agent needs to know what winning looks like in your specific domain. Layer 2 is where that definition lives — including the constraints, business rules, acceptable risk tolerances, and success criteria that distinguish a good outcome from a bad one.
This is distinct from the system prompt. A well-structured goal and context layer includes domain-specific knowledge that the agent can reason against, not just instructions it follows verbatim. CIOs surveyed by CIO magazine in 2026 identified this as one of the most underinvested areas in agentic deployments: organizations focus on the agent's capabilities, but rarely ask whether the agent has been given a coherent model of what success means in their environment.
Without Layer 2, agents produce outputs that look reasonable in isolation but miss the mark against business objectives. They are not hallucinating. They are optimizing for the wrong thing, because no one defined the right thing clearly enough.
This is where the agent's cognitive architecture lives. A planning layer breaks complex tasks into sequential steps, anticipates blockers, and decides what to do next when something goes wrong.
Without it, every multi-step task becomes probabilistic — a coin flip that might land correctly once but fails at scale. Researchers who study production AI agent failures consistently identify the absence of structured planning as a root cause. The pattern is familiar: the model and tools are configured correctly, but the system lacks a decision layer that coordinates them. Each component works. The whole does not.
Good planning architecture separates the reasoning function from the execution function. The reasoning layer decides what to do next given the current context. The execution layer carries it out. Conflating the two — pushing orchestration, retries, and error handling into the same layer as the model's decision logic — is one of the fastest ways to build a system that cannot be debugged under production load.
Agents without persistent memory are stateless. Every interaction starts from scratch. That means no learning from prior sessions, no retention of user preferences or domain context, and no improvement over time.
The result is a system prone to hallucination, inconsistency, and repetition. A well-architected memory layer combines working memory (context maintained within a session), episodic memory (what happened in prior interactions), and semantic memory (the structured knowledge the agent draws on to answer domain-specific questions). Retrieval-augmented generation (RAG) is the most common implementation for semantic memory — connecting the agent to a governed knowledge base that supplements its training data with accurate, current information.
Layer 4 is also where knowledge governance lives. Who updates the knowledge base? How is stale data identified and replaced? What happens when the agent's knowledge contradicts an incoming user instruction? These are not edge cases. They are routine operational questions that need architectural answers before deployment, not after.
This is the layer vendors demo. Update CRMs. Send emails. Spin up workflows. Pull reports. Call APIs. Layer 5 is where the agent acts on the world, and it is genuinely important. But it is also the layer most commonly oversold as sufficient on its own.
Tool use without Layers 1 through 4 produces an agent that can do things but cannot decide wisely which things to do, in what order, or under what conditions. The tools themselves need to be scoped carefully: an agent with broad tool permissions and weak goal context is not powerful — it is a governance risk. Every tool invocation should carry clear permissions, input validation, and error handling. Tool failures cascade to agent failures if the architecture does not account for them explicitly.
Layer 5 is necessary but not sufficient. It is the action layer. The other layers determine whether those actions are correct.
Most production deployments eventually involve more than one agent. A research agent surfaces information. An analysis agent interprets it. A writing agent produces a deliverable. A review agent checks it before output. Each has a narrow scope. Together, they handle a complex workflow.
Without an orchestration layer, this coordination breaks down quickly. Agents produce conflicting outputs. Tasks are duplicated or dropped. Handoffs fail silently. Google DeepMind researchers have identified what they call a "coordination tax" in multi-agent systems: when agents operate without an orchestration layer, coordination overhead grows faster than the productivity gains from adding more agents.
Layer 6 includes the logic that routes tasks between agents, manages state across workflows, handles conflict resolution, and ensures that parallel execution does not produce incoherent results. It is what separates a collection of agents from a functioning system.
The final layer is what separates a pilot from a production system. It includes human approval gates for high-stakes decisions, cost controls that prevent runaway token spend, continuous monitoring that surfaces anomalous behavior before it compounds, and the feedback mechanisms that allow the system to improve over time.
This layer is consistently underbuilt. Research cited across Gartner's 2026 summit series found that fewer than half of organizations have implemented financial guardrails for AI workloads. A separate finding from the same research: 60% of AI projects that lack AI-ready governance will be abandoned before they deliver business value.
Layer 7 is also where the architecture becomes auditable. When something goes wrong — and in production, something always goes wrong — the safety layer is what determines whether you can identify what happened, explain it, and fix it without starting over. Without it, you are not running a production AI system. You are running a pilot that does not know it is a pilot.
The industry's default pattern for AI agent deployment is to start with tool integration and work outward. Connect the CRM. Hook up the email API. Define the workflow. Then ask why the agent is producing off-target outputs and behaving unpredictably under edge cases.
The pattern is understandable. Tool use is visible, demonstrable, and easy to describe in a board presentation. Layers 1 through 4 are invisible until they are absent — and by then, the system is already in production.
The engineering teams that consistently build working agentic systems do the opposite. They define the goal and context layer first. They design the user interaction and feedback loops before writing application code. They build planning and memory before connecting tools. The architecture is defined bottom-up, and tools are added to a system that already knows what it is trying to accomplish.
The seven-layer framework above is technically sound. The harder question is whether your team has the expertise to implement it end-to-end, and whether that expertise exists in-house or needs to be brought in.
Agentic AI architecture is a specialist discipline. It requires strong ML engineering for the reasoning and memory layers, backend architecture skills for orchestration and tool integration, and LLMOps experience for monitoring, governance, and continuous improvement. Building a full-stack team with all of these capabilities in-house is expensive and slow in Western European or North American markets.
This is a significant reason why US, UK, and EU companies are increasingly working with Ukrainian engineering teams for agentic AI development. Ukraine has over 300,000 active software developers — one of Europe's largest technical talent pools — with AI/ML engineering vacancies growing 115% in 2025 alone as demand for agentic systems accelerated. Rates from Ukrainian senior developers typically range from 35–45 EUR/hr, representing savings of 40–60% compared to equivalent hiring in the UK, Germany, or the Netherlands.
What distinguishes Ukrainian AI teams beyond cost is delivery orientation. As ITProfiles' 2026 market analysis notes, Ukrainian engineering teams are characterized by a "bias toward usefulness" — a tendency to integrate AI into business processes early rather than building models in isolation from product outcomes. That bias is directly relevant to the architecture discipline described in this article. Layers 1 and 2 — user interaction and goal context — require engineers who think in terms of business outcomes, not just model performance.
The outstaffing model fits well here. Unlike project outsourcing, outstaffing embeds Ukrainian engineers directly into your team, giving you full control over task allocation, sprint priorities, and architectural decisions. The engineers work within your processes and against your goals. You retain the intellectual property and institutional knowledge. The outstaffing partner handles sourcing, vetting, payroll, and compliance.
For companies building agentic systems, this typically means a hybrid team: your in-house product and tech leads define the architecture and own the product decisions, while Ukrainian engineers implement the layers — with particular strength in the data engineering, RAG, LLMOps, and backend orchestration work that agentic systems require.
Before deploying an AI agent into production, run this checklist against your current architecture. Each question maps to one of the seven layers described above.
A "no" on any of these is not a detail to clean up post-launch. Each missing layer is a specific, predictable failure mode. The architecture either has it or it does not.
The question is not whether AI agents can work. The evidence base for working agentic systems is robust and growing. The question is whether your architecture is built to support them.
Each of the seven layers described in this article is either a point of failure or a competitive advantage. The organizations consistently extracting real business value from AI agents are not the ones with access to better models. They are the ones who invested in the full architecture stack before they shipped — who built the feedback loops, defined the goal context, designed the planning layer, and put safety and governance in place before the first live transaction.
If you are building an agentic system and are not certain every one of these layers is accounted for, that gap is worth addressing before it becomes a production incident.
If you are evaluating options for the engineering capacity to build these systems, working with senior Ukrainian AI engineers through an outstaffing model is one of the most cost-effective and technically credible paths available in 2026. The talent pool is large, the AI specialization is genuine, and the delivery track record speaks for itself.
Ready to build an AI agent architecture that actually works? [Talk to our team about matching you with senior Ukrainian AI engineers — no fluff, just a direct conversation about what you're building and what you need.]
What is AI agent architecture?
AI agent architecture refers to the structural layers that make an autonomous AI system function reliably in production. It includes how the agent receives input, sets goals, plans tasks, retains memory, uses tools, coordinates with other agents, and maintains safety. A well-architected AI agent is more than a model with a prompt — it is a system designed to operate consistently across complex, multi-step tasks.
Why do most AI agent deployments fail?
Most AI agent failures trace back to architectural gaps rather than model limitations. Research from Gartner found that only 23% of failed AI initiatives were caused by model underperformance. The majority fail due to missing governance, weak goal context, absent memory layers, and infrastructure that was not designed for autonomous operation. Organizations that deploy agents without auditing all seven architecture layers are building on a foundation that will not hold under production conditions.
What is the difference between AI agent orchestration and tool use?
Tool use (Layer 5) refers to the agent's ability to call external APIs, update systems, and interact with data sources. Orchestration (Layer 6) refers to the coordination logic that manages how multiple agents or components work together — including task routing, handoffs, state management, and conflict resolution. A single-agent system can function with tool use alone. A multi-agent system requires orchestration to prevent coordination failures.
How does memory work in an AI agent?
Agent memory typically operates across three levels: working memory (context maintained within the current session), episodic memory (records of prior interactions and outcomes), and semantic memory (structured domain knowledge the agent draws on). Retrieval-augmented generation (RAG) is the most common approach for semantic memory, connecting the agent to an external knowledge base that supplements its training data with accurate, current information.
How can I build an AI agent team cost-effectively?
For companies in Western Europe or North America, outstaffing senior Ukrainian AI engineers offers a well-established path to building agentic systems at 40–60% lower cost than local hiring, without sacrificing engineering quality. Ukrainian developers have deep specializations in ML engineering, LLMOps, RAG architecture, and backend orchestration — the exact skill set that agentic system development requires. An outstaffing model gives you full team control while an experienced partner handles sourcing, vetting, and compliance.
Here is a number worth sitting with: 42% of companies abandoned most of their AI initiatives in 2025 — up from just 17% the year before — with an average sunk cost of $7.2 million per failed project. And yet, in the same year, AI investment budgets kept growing. The problem is not enthusiasm. The problem is architecture.
The companies that deploy AI agents and see them fail in production are not using bad models. In most cases, the model performs exactly as advertised. What breaks is everything underneath it — the structural layers that determine whether an agent understands your business, retains context, executes safely, and scales without turning into a liability.
This article breaks down the seven architecture layers that production-ready AI agents require. If you are evaluating whether to build an agentic system, or trying to understand why the one you shipped is underperforming, this is the framework your team needs before writing another line of code.
The industry spent years framing agentic AI as a model selection challenge. Pick the right LLM, engineer a strong prompt, and the agent would figure out the rest. That assumption has now been stress-tested at scale — and it has not held up.
According to Gartner's 2026 analysis, only 23% of failed AI initiatives trace their failure back to model underperformance. The other 77% trace to strategy, governance, and architecture that was never designed to support autonomous systems. A separate Gartner projection warns that 40% of enterprise AI agent projects will be cancelled by 2027 — not because the technology is immature, but because the underlying architecture was not built to support it.
A useful reframe comes from researchers at Halmstad University, who concluded that reliability in agentic systems is "first and foremost, an architectural property." It is not something that emerges from a better model. It emerges from how a system is deliberately decomposed into components, and whether each component is designed to handle real-world load, failure, and ambiguity.
The distinction between an AI chatbot and an AI agent makes this concrete. A chatbot responds to input and completes a single turn. An AI agent plans, acts across multiple steps, loops back on its outputs, and self-corrects when something goes wrong. That is a fundamentally different system — one that needs a fundamentally different infrastructure.
The following framework maps out the structural layers that every production AI agent requires. Every layer is load-bearing. Miss one and you create a system that either fails outright or becomes unpredictable under pressure.
The first layer is the interface between the agent and the humans it serves. This goes beyond the UI. It includes the feedback loops that allow users to set goals, review outputs, and correct the agent's direction over time.
Most teams skip this layer entirely at the design stage. They build the agent's capabilities before defining how users will actually interact with it, what kind of feedback the system will capture, and how that feedback loops back into agent behavior. The result is an agent that operates blind to what the business actually needs — technically functional, strategically useless.
Good user interaction design for agentic systems includes clearly defined input mechanisms, structured output review flows, and explicit checkpoints where a human can redirect or override the agent before it proceeds. Skipping this is not a shortcut. It is a guaranteed path to an agent that diverges from business intent within weeks of deployment.
An agent needs to know what winning looks like in your specific domain. Layer 2 is where that definition lives — including the constraints, business rules, acceptable risk tolerances, and success criteria that distinguish a good outcome from a bad one.
This is distinct from the system prompt. A well-structured goal and context layer includes domain-specific knowledge that the agent can reason against, not just instructions it follows verbatim. CIOs surveyed by CIO magazine in 2026 identified this as one of the most underinvested areas in agentic deployments: organizations focus on the agent's capabilities, but rarely ask whether the agent has been given a coherent model of what success means in their environment.
Without Layer 2, agents produce outputs that look reasonable in isolation but miss the mark against business objectives. They are not hallucinating. They are optimizing for the wrong thing, because no one defined the right thing clearly enough.
This is where the agent's cognitive architecture lives. A planning layer breaks complex tasks into sequential steps, anticipates blockers, and decides what to do next when something goes wrong.
Without it, every multi-step task becomes probabilistic — a coin flip that might land correctly once but fails at scale. Researchers who study production AI agent failures consistently identify the absence of structured planning as a root cause. The pattern is familiar: the model and tools are configured correctly, but the system lacks a decision layer that coordinates them. Each component works. The whole does not.
Good planning architecture separates the reasoning function from the execution function. The reasoning layer decides what to do next given the current context. The execution layer carries it out. Conflating the two — pushing orchestration, retries, and error handling into the same layer as the model's decision logic — is one of the fastest ways to build a system that cannot be debugged under production load.
Agents without persistent memory are stateless. Every interaction starts from scratch. That means no learning from prior sessions, no retention of user preferences or domain context, and no improvement over time.
The result is a system prone to hallucination, inconsistency, and repetition. A well-architected memory layer combines working memory (context maintained within a session), episodic memory (what happened in prior interactions), and semantic memory (the structured knowledge the agent draws on to answer domain-specific questions). Retrieval-augmented generation (RAG) is the most common implementation for semantic memory — connecting the agent to a governed knowledge base that supplements its training data with accurate, current information.
Layer 4 is also where knowledge governance lives. Who updates the knowledge base? How is stale data identified and replaced? What happens when the agent's knowledge contradicts an incoming user instruction? These are not edge cases. They are routine operational questions that need architectural answers before deployment, not after.
This is the layer vendors demo. Update CRMs. Send emails. Spin up workflows. Pull reports. Call APIs. Layer 5 is where the agent acts on the world, and it is genuinely important. But it is also the layer most commonly oversold as sufficient on its own.
Tool use without Layers 1 through 4 produces an agent that can do things but cannot decide wisely which things to do, in what order, or under what conditions. The tools themselves need to be scoped carefully: an agent with broad tool permissions and weak goal context is not powerful — it is a governance risk. Every tool invocation should carry clear permissions, input validation, and error handling. Tool failures cascade to agent failures if the architecture does not account for them explicitly.
Layer 5 is necessary but not sufficient. It is the action layer. The other layers determine whether those actions are correct.
Most production deployments eventually involve more than one agent. A research agent surfaces information. An analysis agent interprets it. A writing agent produces a deliverable. A review agent checks it before output. Each has a narrow scope. Together, they handle a complex workflow.
Without an orchestration layer, this coordination breaks down quickly. Agents produce conflicting outputs. Tasks are duplicated or dropped. Handoffs fail silently. Google DeepMind researchers have identified what they call a "coordination tax" in multi-agent systems: when agents operate without an orchestration layer, coordination overhead grows faster than the productivity gains from adding more agents.
Layer 6 includes the logic that routes tasks between agents, manages state across workflows, handles conflict resolution, and ensures that parallel execution does not produce incoherent results. It is what separates a collection of agents from a functioning system.
The final layer is what separates a pilot from a production system. It includes human approval gates for high-stakes decisions, cost controls that prevent runaway token spend, continuous monitoring that surfaces anomalous behavior before it compounds, and the feedback mechanisms that allow the system to improve over time.
This layer is consistently underbuilt. Research cited across Gartner's 2026 summit series found that fewer than half of organizations have implemented financial guardrails for AI workloads. A separate finding from the same research: 60% of AI projects that lack AI-ready governance will be abandoned before they deliver business value.
Layer 7 is also where the architecture becomes auditable. When something goes wrong — and in production, something always goes wrong — the safety layer is what determines whether you can identify what happened, explain it, and fix it without starting over. Without it, you are not running a production AI system. You are running a pilot that does not know it is a pilot.
The industry's default pattern for AI agent deployment is to start with tool integration and work outward. Connect the CRM. Hook up the email API. Define the workflow. Then ask why the agent is producing off-target outputs and behaving unpredictably under edge cases.
The pattern is understandable. Tool use is visible, demonstrable, and easy to describe in a board presentation. Layers 1 through 4 are invisible until they are absent — and by then, the system is already in production.
The engineering teams that consistently build working agentic systems do the opposite. They define the goal and context layer first. They design the user interaction and feedback loops before writing application code. They build planning and memory before connecting tools. The architecture is defined bottom-up, and tools are added to a system that already knows what it is trying to accomplish.
The seven-layer framework above is technically sound. The harder question is whether your team has the expertise to implement it end-to-end, and whether that expertise exists in-house or needs to be brought in.
Agentic AI architecture is a specialist discipline. It requires strong ML engineering for the reasoning and memory layers, backend architecture skills for orchestration and tool integration, and LLMOps experience for monitoring, governance, and continuous improvement. Building a full-stack team with all of these capabilities in-house is expensive and slow in Western European or North American markets.
This is a significant reason why US, UK, and EU companies are increasingly working with Ukrainian engineering teams for agentic AI development. Ukraine has over 300,000 active software developers — one of Europe's largest technical talent pools — with AI/ML engineering vacancies growing 115% in 2025 alone as demand for agentic systems accelerated. Rates from Ukrainian senior developers typically range from 35–45 EUR/hr, representing savings of 40–60% compared to equivalent hiring in the UK, Germany, or the Netherlands.
What distinguishes Ukrainian AI teams beyond cost is delivery orientation. As ITProfiles' 2026 market analysis notes, Ukrainian engineering teams are characterized by a "bias toward usefulness" — a tendency to integrate AI into business processes early rather than building models in isolation from product outcomes. That bias is directly relevant to the architecture discipline described in this article. Layers 1 and 2 — user interaction and goal context — require engineers who think in terms of business outcomes, not just model performance.
The outstaffing model fits well here. Unlike project outsourcing, outstaffing embeds Ukrainian engineers directly into your team, giving you full control over task allocation, sprint priorities, and architectural decisions. The engineers work within your processes and against your goals. You retain the intellectual property and institutional knowledge. The outstaffing partner handles sourcing, vetting, payroll, and compliance.
For companies building agentic systems, this typically means a hybrid team: your in-house product and tech leads define the architecture and own the product decisions, while Ukrainian engineers implement the layers — with particular strength in the data engineering, RAG, LLMOps, and backend orchestration work that agentic systems require.
Before deploying an AI agent into production, run this checklist against your current architecture. Each question maps to one of the seven layers described above.
A "no" on any of these is not a detail to clean up post-launch. Each missing layer is a specific, predictable failure mode. The architecture either has it or it does not.
The question is not whether AI agents can work. The evidence base for working agentic systems is robust and growing. The question is whether your architecture is built to support them.
Each of the seven layers described in this article is either a point of failure or a competitive advantage. The organizations consistently extracting real business value from AI agents are not the ones with access to better models. They are the ones who invested in the full architecture stack before they shipped — who built the feedback loops, defined the goal context, designed the planning layer, and put safety and governance in place before the first live transaction.
If you are building an agentic system and are not certain every one of these layers is accounted for, that gap is worth addressing before it becomes a production incident.
If you are evaluating options for the engineering capacity to build these systems, working with senior Ukrainian AI engineers through an outstaffing model is one of the most cost-effective and technically credible paths available in 2026. The talent pool is large, the AI specialization is genuine, and the delivery track record speaks for itself.
Ready to build an AI agent architecture that actually works? [Talk to our team about matching you with senior Ukrainian AI engineers — no fluff, just a direct conversation about what you're building and what you need.]
What is AI agent architecture?
AI agent architecture refers to the structural layers that make an autonomous AI system function reliably in production. It includes how the agent receives input, sets goals, plans tasks, retains memory, uses tools, coordinates with other agents, and maintains safety. A well-architected AI agent is more than a model with a prompt — it is a system designed to operate consistently across complex, multi-step tasks.
Why do most AI agent deployments fail?
Most AI agent failures trace back to architectural gaps rather than model limitations. Research from Gartner found that only 23% of failed AI initiatives were caused by model underperformance. The majority fail due to missing governance, weak goal context, absent memory layers, and infrastructure that was not designed for autonomous operation. Organizations that deploy agents without auditing all seven architecture layers are building on a foundation that will not hold under production conditions.
What is the difference between AI agent orchestration and tool use?
Tool use (Layer 5) refers to the agent's ability to call external APIs, update systems, and interact with data sources. Orchestration (Layer 6) refers to the coordination logic that manages how multiple agents or components work together — including task routing, handoffs, state management, and conflict resolution. A single-agent system can function with tool use alone. A multi-agent system requires orchestration to prevent coordination failures.
How does memory work in an AI agent?
Agent memory typically operates across three levels: working memory (context maintained within the current session), episodic memory (records of prior interactions and outcomes), and semantic memory (structured domain knowledge the agent draws on). Retrieval-augmented generation (RAG) is the most common approach for semantic memory, connecting the agent to an external knowledge base that supplements its training data with accurate, current information.
How can I build an AI agent team cost-effectively?
For companies in Western Europe or North America, outstaffing senior Ukrainian AI engineers offers a well-established path to building agentic systems at 40–60% lower cost than local hiring, without sacrificing engineering quality. Ukrainian developers have deep specializations in ML engineering, LLMOps, RAG architecture, and backend orchestration — the exact skill set that agentic system development requires. An outstaffing model gives you full team control while an experienced partner handles sourcing, vetting, and compliance.
.svg)
.svg)
.svg)
