For the five implementation studies and frontier research, start with the Agent Deep Dives overview. This directory preserves the broader set of products, harnesses and development frameworks.
Agent products can look similar while owning very different parts of the system. A coding application, an SDK and a protocol are not interchangeable alternatives. Compare the responsibility you need to keep or delegate before comparing names.
Directory source check: October 10, 2026. The five linked deep dives have their own October 11 checks. This is a selected architectural map, not an exhaustive catalog or performance ranking. Official descriptions establish documented behavior, not independent evidence of reliability. Features and availability can vary by edition, configuration and account.
First identify the layer
| Layer | Main responsibility | Example question |
|---|---|---|
| Model | Interpret context and propose decisions | Which capabilities does this task require? |
| Harness | Run the loop, tools, context and stopping rules | Who owns recovery and policy enforcement? |
| Product | Provide an end-user workflow | Is work interactive, delegated or embedded? |
| Framework or SDK | Supply reusable implementation primitives | Which runtime responsibilities remain ours? |
| Protocol | Connect components through a shared contract | Which endpoints and extensions interoperate? |
This is our comparison taxonomy. A project can span multiple rows, and a managed product may contain several of these layers internally.
Coding products and toolkits
The architecture descriptions below summarize inspected official sources. The questions in the final column are our suggested review criteria.
| Project | Documented work shape | Inspect before adoption |
|---|---|---|
| Claude Code | Coding harness with context gathering, actions, verification and multiple execution interfaces.1 | Execution location, tool access and review boundaries |
| Codex | Local integration through the SDK and separate cloud task workflows with reviewable changes.23 | Which interface owns the session, environment and approvals |
| GitHub Copilot cloud agent | Repository work in an ephemeral cloud development environment.4 | Repository access, workflow integration and review requirements |
| OpenCode | Coding application with terminal, desktop and editor surfaces and configurable providers.5 | Provider setup, permissions and actual runtime configuration |
| Pi | Modular toolkit containing model APIs, agent runtime and coding CLI.6 | External containment: the repository states there is no built-in permission system |
A terminal interface does not prove that all execution is local. A cloud interface does not establish that your data and credentials have appropriate boundaries. Read the execution and security contracts, not only the screenshot.
Reusable harnesses and frameworks
| Project | Documented design focus | Important distinction |
|---|---|---|
| Deep Agents | Reusable harness with filesystem, context management, delegation and optional task planning.7 | Built on LangChain primitives and the LangGraph runtime, not a separate model |
| DeepSeek Harness | Plugin-oriented agent harness.8 | Still a developer preview with explicit compatibility-breaking warnings |
| smolagents | Python agents whose actions can be expressed as code, plus conventional tool-calling agents.9 | Code actions are not limited to writing software; execution still needs containment |
| LangGraph | Stateful execution with thread checkpoints and separate cross-thread stores.10 | Persistence primitives do not automatically make external effects safe to replay |
| OpenAI Agents SDK | Application-owned agent harness with sandbox integration.11 | Different ownership from the managed Agents API |
| Microsoft Agent Framework | Agents, state, middleware and graph-based workflows; successor direction for AutoGen and Semantic Kernel.12 | Check feature support for the language and package edition you use |
| Google ADK | Tools for agent development, graph and collaborative workflows, evaluation and deployment.13 | Inspect the chosen language and deployment path rather than infer support from the umbrella name |
| Strands Harness SDK | Agent loop, providers, tools and orchestration primitives in a Python and TypeScript monorepo, alongside an assembled harness.14 | Distinguish the assembled harness from the SDK used to construct one |
These are useful study candidates, not recommendations to adopt every framework. A small explicit loop can be easier to inspect than an abstraction whose persistence, retry or approval semantics you do not understand.
Managed runtime ownership
OpenAI’s Agents API hosts the Codex harness while the application submits work and handles its integration. It can use hosted compute, connected compute or no execution environment. Anthropic’s managed architecture separates the session, harness and sandbox so their lifecycles can differ.1516
The key distinction is who owns each responsibility, not whether a service is marketed as an agent platform. Compare context handling, persistent state, tool execution, cancellation, approvals, observability and cleanup. Managed does not mean that the application has no security or correctness work left.
Community developments worth understanding
The following are themes supported by the selected sources, not a claim that each was invented recently:
- Procedure packaging: interoperable skills separate reusable instructions from callable tools. See Tools, Skills & Protocols.
- Selective context: tool discovery and programmatic processing reduce what must enter each model call. The tradeoffs depend on the workload.
- Durable execution: sessions, checkpoints and artifacts support work beyond one process. See Context, Memory & State.
- Explicit boundaries: managed control and execution environments can have separate ownership and lifecycles.
- Repeatable evaluation: task and grader editions, repeated trials and budgets belong beside results. See Evaluation & Reliability.
How this map should evolve
For future revisions, recheck the underlying official source before updating a product claim. Record meaningful renames, deprecations, ownership changes and security defaults. Advance the source-check date only after the mutable comparisons have been rechecked.
A reasonable editorial practice is to review this map monthly or before a technology decision, while revising the foundations when understanding changes. This is a suggested cadence, not an automated monitoring service.
References
-
Anthropic. How Claude Code works. Checked October 10, 2026. ↩
-
OpenAI. Codex Cloud. Checked October 10, 2026. ↩
-
GitHub. Copilot on GitHub.com. Checked October 10, 2026. ↩
-
OpenCode. Documentation introduction. Checked October 10, 2026. ↩
-
Earendil Works. Pi repository, packages and Permissions & Containerization. Checked October 10, 2026. ↩
-
LangChain. Deep Agents overview, capabilities and optional planning. Checked October 10, 2026. ↩
-
DeepSeek AI. DeepSeek Harness repository, developer preview notice. Checked October 10, 2026. ↩
-
Hugging Face. smolagents introduction, CodeAgent and ToolCallingAgent. Checked October 10, 2026. ↩
-
LangChain. LangGraph persistence. Checked October 10, 2026. ↩
-
OpenAI. Sandbox Agents, application-owned SDK harness. Checked October 10, 2026. ↩
-
Microsoft. Agent Framework overview, workflows and relationship to AutoGen and Semantic Kernel. Checked October 10, 2026. ↩
-
Google. Agent Development Kit. Checked October 10, 2026. ↩
-
Strands Agents. Harness SDK monorepo, current destination of the earlier Python SDK repository. Checked October 10, 2026. ↩
-
OpenAI. Agents API architecture. Checked October 10, 2026. ↩
-
Anthropic. Scaling Managed Agents: Decoupling the brain from the hands, April 8, 2026. ↩