Ecosystem-specific
Materially tied to one vendor's managed platform, hosted service, API or product ecosystem. Often the right engineering choice when that ecosystem already fits your system.
ECOSYSTEM · WHERE COGNOKRATOS FITS
Learn the architecture. Then choose the implementation.
A short, curated guide to the courses, frameworks, runtimes, policy engines, standards and infrastructure around CognoKratos. This page orients; the book chapter explains.
01 / THE MAP
Between the fundamentals and your system.
CognoKratos studies where trust comes from when these components are composed into systems whose actions matter.
The projects on this page solve problems CognoKratos deliberately does not try to own. The curriculum focuses on the boundaries between them: identity, capability, policy, durability, authority, evidence, custody and settlement. These are a working vocabulary, not a standard.
Inclusion is not endorsement, and no project here is a dependency of CognoKratos. Descriptions follow each project’s documentation as checked in October 2026. Projects change, so read their current documentation before relying on them.
Learn the fundamentals
- Hugging Face
- Microsoft
- AWS
- Anthropic
- OpenAI
CognoKratos
Learn the architecture and trust boundaries
- Identity
- Capability
- Policy
- Durability
- Authority
- Evidence
- Custody
- Settlement
Frameworks & runtimes
- NeMo Agent Toolkit
- Strands
- OpenAI Agents SDK
- Dapr Agents
- Rig
Governance & policy
- Agent Governance Toolkit
- AgentCore Policy
- Agent Control Standard
- OPA
- Cedar
- OpenFGA
Protocols & infrastructure
- MCP
- ERC-8004
- AgentKit
- x402
- Nevermined
Build your system
02 / THREE LABELS
Coupling, not quality.
Each external project carries one label answering a single question: how tightly is it coupled to one vendor or platform? It says nothing about which is better, and learning paths are described by their ecosystem instead.
Portable open source
Open-source software you can run yourself, usable across providers or environments, even when one company leads its development. The label is about portability, not neutral governance.
Open standard / infrastructure
Open standards, specifications, protocols and foundation- or community-governed infrastructure intended to work across vendors and stacks: protocols such as MCP, OIDC or x402, and components such as OPA, Cedar or OpenFGA. Governance and maturity differ, and some are draft or pre-1.0.
03 / 5 CURATED RESOURCES
Need stronger fundamentals?
CognoKratos assumes you can already build a small tool-using agent. If not, choose the path that matches the ecosystem you already work in; none is the right one. CognoKratos begins where they usually stop: when execution, authority, failure and evidence begin to matter.
- Multi-framework path Hugging Face Agents Course
A free course on agent fundamentals, shown in smolagents, LlamaIndex and LangGraph. Seeing one idea in several frameworks helps separate concepts from implementations. Maintained by Hugging Face, whose inference it uses by default.
USE WHENBook · If you are new to agents ↗you have no particular ecosystem in mind and want the broadest foundation.
- Microsoft-focused Microsoft AI Agents for Beginners
Eighteen open lessons on agent patterns, tool use, RAG, multi-agent systems, memory and security. Its exercises use Microsoft Agent Framework with Microsoft Foundry.
USE WHENyou build on Microsoft’s platform.
- AWS-focused Getting Started with Strands Agents
A hands-on AWS learning path: build a first agent with Strands, then tools and MCP, multi-agent patterns, memory, observability and deployment on Amazon Bedrock AgentCore.
USE WHENyou want to learn agents through the AWS ecosystem.
- Anthropic-focused Claude Academy
Anthropic’s maintained learning platform for working with Claude. Its Claude Platform courses cover API fundamentals, building with the Claude API (tool use, retrieval, agents) and the Model Context Protocol. The Claude Cookbooks and tool-use docs make good further reading.
USE WHENyou primarily work with Claude.
- OpenAI-focused OpenAI Cookbook
Worked examples for the OpenAI platform, including agents built with the open-source OpenAI Agents SDK.
USE WHENyou build on OpenAI’s platform.
04 / 10 CURATED RESOURCES
Need production implementations?
Frameworks, durable runtimes and governance tooling that implement concepts the projects teach. Each makes some properties easier to obtain; none removes the need to know which ones you depend on.
- Portable OSS NVIDIA NeMo Agent Toolkit
Configuration, MCP, observability, evaluation and optimisation for agents built with many frameworks. Does not require NVIDIA hardware.
USE WHENPart I · Production agents ↗you want a mature toolkit around your agents. It runs the canonical simple-agent-template.
- Portable OSS Strands Agents
An open-source agent SDK from AWS for Python and TypeScript. Runs in your process, with Amazon Bedrock as the default among many model providers.
USE WHENPart I · Production agents ↗you want a lightweight SDK that fits naturally on AWS without requiring it.
- Portable OSS OpenAI Agents SDK
A small framework for agents, handoffs, guardrails, human-in-the-loop and tracing. Works with other model providers; its hosted tools and default trace export use the OpenAI platform.
USE WHENPart I · Production agents ↗you want minimal primitives and may mix providers.
- Portable OSS Rig
A Rust library for LLM applications with many model providers behind one interface. Pre-1.0.
USE WHENPart I · Production agents ↗you want to own the agent loop in Rust. It runs the Rust alternative of simple-agent-template.
- Ecosystem-specific Amazon Bedrock AgentCore
AWS’s managed platform for running and operating agents built with any framework and model, with services such as runtime, gateway, identity, memory, observability and policy.
USE WHENPart I · Production agents ↗you want AWS to operate the agent runtime and its surrounding services.
- Portable OSS Dapr Agents
Python agents whose execution runs on Dapr Workflows, from the CNCF Dapr project.
USE WHENPart II · Durable runtimes ↗agents must survive restarts and you already run, or want, Dapr.
- Portable OSS Temporal
A durable execution platform with integrations for agent frameworks. Activities can be retried, so side effects still need idempotency.
USE WHENPart II · Durable runtimes ↗long-running agent workflows need durable execution at scale.
- Portable OSS Microsoft Agent Governance Toolkit
Policy enforcement, identity, sandboxing and audit around agent actions, with adapters for many frameworks. Does not require Azure. In public preview.
USE WHENPart III · Governed decisions ↗you need runtime governance across agent frameworks.
- Ecosystem-specific AgentCore Policy
AWS-managed policy enforcement: Cedar policies evaluated on agent tool calls that pass through an AgentCore Gateway, with each decision logged. The same boundary as the toolkit above, as a managed service you configure.
USE WHENPart III · Governed decisions ↗your agents already run on Amazon Bedrock AgentCore.
- Open standard & infra OPA, Cedar and OpenFGA
Policy and authorisation engines, all CNCF projects at different maturity levels: OPA with Rego, Cedar (originated at AWS) with analysable policies, OpenFGA with relationship-based access control.
USE WHENPart III · Governed decisions ↗permission decisions should live in reviewed policy rather than application branches.
05 / 8 CURATED RESOURCES
Need standards or infrastructure?
Protocols, identity standards and economic infrastructure. Blockchain appears here as one possible layer for identity, custody and settlement, not as a default.
- Open standard & infra Model Context Protocol
The tool protocol used across all five projects, now a project of the Linux Foundation’s Agentic AI Foundation. A protocol, not a security model.
USE WHENPart I · Production agents ↗tools should sit behind a process and capability boundary.
- Open standard & infra Agent Control Standard
An early OWASP GenAI Security Project specification for a separate guardian that can permit, deny or modify agent actions.
USE WHENPart III · Governed decisions ↗you are designing control points outside the agent itself.
- Open standard & infra OpenID Connect and SPIFFE
Identity for people (OIDC) and for workloads (SPIFFE, a CNCF standard).
USE WHENPart I · Production agents ↗you must establish who, or what, is calling before anything else.
- Open standard & infra OpenTelemetry GenAI conventions
Shared names for model, agent and tool spans. Still in development.
USE WHENPart I · Production agents ↗traces must stay readable across frameworks and vendors.
- Portable OSS Coinbase AgentKit
Wallets and on-chain actions for agents, through pluggable wallet providers ranging from Coinbase-managed wallets to keys you hold.
USE WHENPart IV · Cryptographic capability ↗an agent needs an on-chain capability and you must choose who holds the key.
- Open standard & infra ERC-8004: Trustless Agents
A draft Ethereum standard for agent identity, reputation and validation registries. Payments are out of its scope.
USE WHENPart IV · Cryptographic capability ↗agents must be discoverable and identifiable across organisations.
- Open standard & infra x402
An open protocol for payments over HTTP, stewarded by the x402 Foundation under the Linux Foundation.
USE WHENPart V · Financial workflows ↗an agent or service must pay or charge per request.
- Ecosystem-specific Nevermined
A platform for metering, access control and payments between agents, with open-source SDKs.
USE WHENPart V · Financial workflows ↗you want a managed path to charging for agent services.
06 / ONE LESSON · TWO IMPLEMENTATIONS
Replace the framework. Keep the architecture.
simple-agent-template is built twice, so you can test whether you still recognise
the architecture after the framework is replaced.
On main it runs on NVIDIA NeMo Agent Toolkit, a mature framework that
provides most behaviour through configuration. On rust-agent the same agent
service is written in Rust on Rig, which leaves most of that behaviour to ordinary code.
The gateway, MCP server, database, prompts, contract and evaluation suites are shared.
The two are not at full parity, and the differences are documented: PII masking is narrower on Rig, tool-argument checks are stricter, and NeMo Agent Toolkit remains the recommended production reference. Neither is presented as better.
Same engineering lesson
Identity · capability · authority · evidence
NeMo Agent Toolkit
Python · main
Rust + Rig
rust-agent
Same architectural properties
One gateway, one MCP server, one contract, one evaluation
07 / THE IMPLEMENTATION SPECTRUM
Delegate or own. Choose knowingly.
Every trust boundary can be built at a different distance from your own code. The right position depends on your risk model, operations, compliance, portability, expertise, economics and trust boundaries.
- 01
Managed platform
More implementation and operation delegated to the provider. Some controls become properties of the service and its configuration, which you still configure and verify.
Amazon Bedrock AgentCore, Microsoft Foundry
- 02
Portable open-source framework
Code you run and can read, with a company setting its direction.
NeMo Agent Toolkit, Agent Governance Toolkit, Strands, OpenAI Agents SDK
- 03
Open standards and components
Standards and engines you compose yourself, under open governance.
MCP, Cedar, OPA, OpenFGA
- 04
Minimal or custom implementation
The most owned: every property visible in your code, and your responsibility.
Rust on Rig, application-owned controls
The lesson is not that the bottom is better, or that managed means guaranteed. A managed platform moves trust boundaries and takes on part of the work, but configuring, integrating and verifying your side remains yours. One system can sit at different points for different boundaries: agents on a managed platform, policy in open Cedar, approvals checked in its own database transaction. What matters is knowing which responsibility you have delegated, to whom, and how you would notice if it stopped holding.
ABOUT THESE REFERENCES
CognoKratos references external projects because good engineering starts by understanding and reusing existing work. A resource is listed because it illustrates an implementation, a standard or an alternative relevant to a problem the laboratories study, and the list is short on purpose. The book chapter Where CognoKratos fits covers more projects and maps them to trust boundaries. Found a description that is out of date? Tell us how to correct it.