Context & tools
The material the work starts from: the sources, documents and tools a team already relies on.
Updated May 15, 2026 · 7 min read
AI Agent Tools (September 2026)
The 2026 agent stack has stabilised into roughly 12 layers: models, gateways, frameworks, runtimes, no-code builders, coding agents, MCP tool registries, eval, memory, voice, sandboxed execution, and browser / computer use. This is the category-by-category breakdown of what to pick at each layer.
Twelve layers of the AI agent stack
Frontier model providers
Anthropic (Claude Opus 4.7, Sonnet 4, Haiku 3.5), OpenAI (GPT-5.5 Pro, GPT-5.5, mini, nano), Google (Gemini 3.1 Pro, 3.0, 2.5 Flash), DeepSeek V4, xAI Grok 4. The model layer, pick by workload, route via gateway.
AI gateways + LLM proxies
Swfte, OpenRouter, Portkey, LiteLLM, TrueFoundry, EdenAI, Cloudflare AI Gateway. Route across providers, enforce policy, cache prompts, attribute cost.
Agent frameworks
LangGraph (LangChain), CrewAI, AutoGen, Letta, OpenAI Agents SDK, Anthropic Computer Use, Agno. Build agents in code with explicit state, tools, and orchestration.
Managed agent runtimes
Swfte, Vellum (post-pivot), Relevance AI, Glean, StackAI, Dust. The runtime layer, host the agents, eval them, govern them, attribute cost.
No-code agent builders
Gumloop, Lindy, Botpress, MindStudio, Relay, FlowiseAI. Visual canvas for operators and non-technical users.
Coding agents
Claude Code, Cursor, Cline, Aider, Continue, OpenCode, GitHub Copilot. Agents that read, write, and ship code.
Tool registries + MCP servers
Anthropic MCP, Pulse MCP Hub, mcp-servers community catalog, Swfte MCP gateway. Standardised way to expose tools to agents.
Eval + observability
Swfte built-in, LangSmith, Langfuse, Arize Phoenix, Galileo, Patronus, Promptfoo. Measure quality, catch regressions, run shadow A/Bs.
Memory + retrieval
Pinecone, Weaviate, Qdrant, pgvector, MongoDB Atlas Vector, LlamaIndex, Letta memory. Long-term context for agents that need to remember.
Voice + multimodal
OpenAI Realtime, Gemini Live, ElevenLabs, Cartesia, Deepgram, Whisper, Recall.ai. Voice + multimodal tooling for agents that talk and listen.
Sandboxed execution
E2B, Modal, Daytona, Vercel Sandbox, AWS Lambda. Run agent-generated code safely.
Browser + computer use
Anthropic Computer Use, OpenAI Operator, Browserbase, Playwright + LLM wrappers. Agents that drive a browser or full desktop.
How to assemble agent tools by stage
| Stage | Tool picks + advice |
|---|---|
| Stage 1: Prototype | Pick a frontier model (Claude Sonnet 4 default) + an agent framework (CrewAI or LangGraph) + a sandbox (E2B). Skip the gateway, skip the eval harness. Ship one working agent. |
| Stage 2: Production | Add a gateway (Swfte or LiteLLM) for routing + caching + cost ceilings. Add an eval harness (Swfte built-in, LangSmith, or Langfuse). Add observability traces. |
| Stage 3: Fleet | Move to a managed runtime (Swfte) that bundles gateway + agent runtime + eval + observability on one platform. Add per-team budgets, audit log, on-prem option if regulated. |
| Stage 4: Org-wide | Add an MCP gateway in front of the tool catalog. Add an AI governance program (NIST AI RMF + EU AI Act mapping). Promote model owners and per-agent KPIs. |
FAQ
What are AI agent tools?
AI agent tools are the software building blocks that make modern agents work: the model providers, the gateway that routes between them, the framework or runtime that orchestrates steps, the eval harness, the tool registry, the memory layer, and the execution sandbox. By 2026 the stack has stabilised into roughly 10 categories.
What is the difference between agent frameworks and agent runtimes?
Agent frameworks (LangGraph, CrewAI, AutoGen) are code-level libraries: you import them, write your agent logic, and run it on infrastructure you operate. Agent runtimes (Swfte, Vellum, Relevance AI) are managed platforms: you describe your agent and the runtime handles the orchestration, scaling, retries, eval, and governance. Frameworks suit prototyping; runtimes suit production fleets.
Do I need an agent framework + a runtime?
Depends. Many production agents run on a runtime alone (Swfte definitions describe the agent declaratively, no framework needed). Teams that want bespoke control may still use a framework on top of a runtime. The mature 2026 pattern is to keep the runtime managed and use a framework only when a specific orchestration shape demands it.
Which agent tools should I pick first?
Three picks cover 90% of starter agents. A frontier model: Claude Sonnet 4. An orchestration layer: a framework like CrewAI or a runtime like Swfte. A sandbox for code-running agents: E2B or Modal. Add a gateway and eval harness once you have a second agent in production.
What is an MCP server?
An MCP (Model Context Protocol) server exposes tools and resources to AI agents through a standardised wire protocol. Once a tool is exposed as an MCP server, any MCP-aware model (Claude, GPT, Gemini, DeepSeek, Grok) can discover and call it without bespoke integration code. MCP is the closest the agent ecosystem has to a tool-call standard.
Best agent tools for coding agents?
Claude Code or Cursor as the agent surface, Claude Opus 4.7 as the underlying model (1567 Elo on the coding leaderboard), E2B or Modal as the code sandbox, Swfte as the gateway for cost ceilings + audit. Most senior engineers run two surfaces at once: Cursor for interactive editing, Claude Code for terminal-native automation.
Best agent tools for voice agents?
OpenAI Realtime as the voice model, Twilio or Bland AI as the phone number / call infrastructure, Vapi or Retell AI as the agent orchestration, Swfte voice + chat template for the runtime, Recall.ai if the workflow includes meeting capture.
How do I evaluate AI agent tools?
Three KPIs. Task completion rate on a golden dataset (does the agent finish the task correctly). Cost per successful completion (tokens + tool calls + downstream API costs). Safety (does the agent stay within policy on every step). Swfte's built-in eval harness, LangSmith, and Langfuse all give you these. Set up the eval harness before scaling the agent.
Are AI agent tools mature in 2026?
For coding and customer support, yes. Mature, production-grade tools exist at every layer. For multi-agent autonomous systems and computer-use agents, the stack is functional but still moving: expect 1-2 framework upgrades per quarter and ongoing API changes from the model providers. Plan operating cost for the upgrade lane.
Run the full agent stack on one runtime
Swfte bundles the gateway, the agent runtime, the eval harness, the MCP gateway, and per-team cost ceilings on a SOC2 platform.
Free tier · OpenAI-compatible API · SOC2 Type II · On-prem available
Workflow study / Agent Tools
From prototype tools to a reviewed release
Choose the stage to see which engineering question comes next. Illustrative authored example · snapshot · 7 September 2026. Supplied catalogue copy, not verified performance evidence; this preview does not execute a workflow.
Input
Build an agent that reads a support ticket and suggests a reply.
Start with a model, task runner and a small representative example set.
Selection question: can the agent complete the intended task?
Input
The agent now needs access to the ticketing system.
Add authenticated tool access, a runtime boundary and traces.
Selection question: can the team explain and constrain its actions?
Input
Several teams want to operate their own agents.
Introduce shared policy, ownership and evaluation records.
Selection question: who owns each agent and its release decision?
How the pieces connect
One useful outcome.
Six deliberate layers.
A spatial explanation of the workflow.
Select a stage or scroll to explore.
EXPLORE THE ASSEMBLY
Intelligence
The reasoning step. A model reads the context that was gathered and proposes something a person can act on.
Review boundaries
The limits set before the work runs — what is in scope, what is refused, and who is asked when it is unclear.
Workflow
The sequence itself: the order of steps, the handoffs between them, and where a person is required.
Deployment
How the result reaches the people who use it, and the conditions under which it is allowed to run.
Evidence
What is kept afterwards so a reviewer can retrace the decision: inputs, choices, and the person accountable.
CONCEPTUAL WORKFLOW MODEL — NOT A DEPICTION OF SYSTEM ARCHITECTURE
Illustrative authored example · snapshot · 7 September 2026. Capabilities and figures require confirmation.