Build Agents
Platforms
Current Anthropic, OpenAI, and Gemini platform surfaces for agent systems - hosted tools, MCP connectors, Agent Skills, code execution, and model guidance.
Last verified 2026-10-29
Summary
Treat Claude, OpenAI, and Gemini as platforms: all three ship a hosted execution surface, and Claude and OpenAI add connectors for reaching external tools plus a shared cross-vendor Agent Skills mechanism. The common failure is scaffolding against last quarter's shape - a retired , a superseded model ID, or a renamed parameter.
Anthropic docs live at platform.claude.com; Claude Code docs live at code.claude.com. OpenAI docs live at developers.openai.com (the old platform.openai.com paths redirect there). Gemini docs live at ai.google.dev and blog.google.
Anthropic
Choose Managed Agents when sessions, credential vaulting, , or scheduled runs should be Anthropic-operated rather than embedded in your own app server. Choose Claude Code / Agent SDK when a human or service needs the coding harness directly. Choose the MCP connector when a single Messages API request just needs remote tool calls. Choose a custom MCP server when Claude, OpenAI, and other clients all need the same vendor-neutral contract.
Current model IDs as of 2026-10-29: Opus 5.5 (claude-opus-5-5) is the recommended default for agent and coding work; Fable 5.1 (claude-fable-5-1) is the flagship for deep reasoning and long tasks; Sonnet 5.5 (claude-sonnet-5-5, 2026-09-28) is the balanced mid tier; Haiku 5.5 (claude-haiku-5-5, 2026-10-07) is the low-latency option. Current pinned IDs are dateless - do not append a date suffix. See Model Reference for the full allowlist.
Anthropic API version 2026-10-21 is the default for the new major client SDKs. It removes temperature, top_p, and top_k (a request that sends them returns a 400). The Claude Agent SDK is versioned separately.
OpenAI
Any code still targeting the Assistants API belongs on the migration guide below; build new work directly on Responses + Conversations. Choose the Agents API when the harness itself - orchestration, compaction, recovery - should be Codex-operated rather than assembled from the SDK; choose the Agents SDK when the app needs to own that loop directly.
Current model IDs as of 2026-10-29: GPT-6 Astra (gpt-6-astra, flagship), GPT-6.1 Sol (gpt-6.1-sol), GPT-6.1 Terra (gpt-6.1-terra, balanced default, 2026-10-26), GPT-6 Luna (gpt-6-luna, fastest/cheapest). See Model Reference.
Gemini
Google's managed-agent surface on the Gemini API, expanded 2026-07-28:
| Surface | Use when | Notes |
|---|---|---|
| Environment hooks | Custom checks should run around every tool call | A .agents/hooks.json script runs on pre_tool_execution and post_tool_execution inside the sandbox - block, lint, or audit a tool call before or after it executes. |
| Token budgets | A run must not exceed a known cost | max_total_tokens caps input, output, and thinking combined; the task pauses safely at the limit and returns status: "incomplete" with environment state preserved, rather than erroring out. |
| Cron scheduling | An agent should run unattended on a schedule | Triggers bind an agent, environment, prompt, and cron schedule into a persistent resource that fires without manual intervention; sandboxes persist across scheduled runs. |
| Environments API | Sandbox sessions need programmatic lifecycle management | List, inspect, and delete environments directly instead of waiting on the default 7-day timeout. |
Gemini 3.8 Flash (gemini-3.8-flash) is the Flash model Google's catalogue now recommends for new work, with Gemini 3.5 Flash-Lite for cost-sensitive runs; check managed-agent endpoint support before changing a deployment pin from 3.7 Flash. The hook, budget, scheduling, and Environments API details above come from Google's July 2026 announcement, and no dated managed-agent API reference was found on 2026-10-29 to re-confirm them. Google's published docs do not describe an Agent Skills mechanism or an MCP connector for the managed-agent path, so confirm both before relying on them. Treat this as a narrower surface than Claude Managed Agents or the OpenAI Agents API, and treat it as the right fit for scheduled, budget-bounded background agents on Gemini rather than a general replacement for the other two.
MCP Notes
None of the three platforms support the full MCP primitive set through every path. State which path is in use, and do not assume support for resources, prompts, roots, sampling, elicitation, or tasks unless the specific surface documents it:
- Claude Code uses local stdio and remote Streamable HTTP MCP servers via
.mcp.json, and can itself act as an MCP server withclaude mcp serve. - The Claude Messages API MCP connector supports
mcp_toolsetwithdefer_loadingfor large catalogs and , but only MCP tool calls. - OpenAI can call remote MCP servers as hosted tools from a Responses request, plus OpenAI-maintained connectors for common third-party systems; private servers reach it through a secure tunnel.
See MCP Protocol and MCP Auth Model for the protocol-level detail these platforms build on.
Security Guidance
The same rules apply across all three platforms:
- Treat model-visible tool descriptions, , connector metadata, and Skill content as advisory input, never a trust boundary.
- Enforce authorization, tenant identity, and checks in trusted server code - not in the prompt.
- Gate destructive, external, authenticated, production, browser, and sandbox/container tools with confirmation policies (see Prompts and Configuration and Guardrails).
- Keep production credentials out of prompts, Skill text, and sandbox/container environments; broker privileged actions through a server-side tool the sandbox cannot bypass.
- Log agent/session identity, user or tenant context, tool calls, approvals, and outcomes.