Bound the job.
“Resolve this class of invoice discrepancy” is a useful goal. Define the inputs, completion criteria and actions an agent may take.
Beyond the chatbot.
Meet the agents.
Understand the model, the harness and the workflow around it. Each layer brings different strengths, costs and responsibilities.
5 harness profiles·8 model assessments·Reviewed 20 September 2026
MAKE AN INFORMED CHOICE
Which model fits the work?
Specific models, their advantages, their limitations and the jobs worth testing them on.
Compare models NEW IDEAS, EXAMINEDJev and what comes next.
Fast decision models, managed harnesses and evolving ways to organise agent work.
Explore innovation watchA model can think.
A harness helps it work.
A model generates decisions or language. An agent uses a model to choose and execute steps towards a goal. Its harness supplies the execution loop, tools, context, memory, permissions and records. An orchestration framework connects those pieces into a process. A model comparison and a framework comparison answer different questions.
Human intent
Goal · limits · success criteria
The agent loop
Plan · act · observe · adapt
Tools & systems
Data · software · actions
Verified outcome
Evidence · review · learning
The harness surrounds the loop with permissions, context management, evaluation and recovery. Humans remain accountable for the process.
Make judgement explicit.
Give people evidence and authority to review consequential decisions. Connect approvals to execution permissions and provide a clear escalation route.
Build for recovery.
Record actions, detect repeated attempts and test partial failures. A useful agent can stop safely and help a person understand what happened.
THE AGENTIC STACK
Choose the right foundation.
Different levels of control, different operational responsibilities.
LangGraph
Explicit control over long-running, stateful workflows.
Best fit: Teams combining predictable business rules with agent decisions, checkpoints and human review.
- Persistent state and recovery
- Human review through interrupts
- Mix deterministic and model-driven steps
The trade-off: You design more of the architecture. Deep Agents adds a higher-level harness on top when you want planning, filesystem tools and subagents.
Read the documentationOpenAI Agents SDK
Tools, delegation and observability in a compact runtime.
Best fit: Developers building tool-using applications, specialist handoffs and agents that work inside isolated workspaces.
- Built-in agent loop and tracing
- Handoffs, tools and guardrails
- Sessions and sandbox agents
The trade-off: A library still needs an application and deployment design. Distinguish it from the separately managed Agents API; your runtime choice affects operational ownership.
Read the documentationClaude Agent SDK
Claude Code's agent loop, tools and context management as a library.
Best fit: Python and TypeScript applications that need agents to inspect files, execute commands and complete multi-step work.
- Built-in file and command tools
- Permissions, hooks and sessions
- MCP, skills and subagents
The trade-off: The execution environment and tool permissions need careful design. The SDK and Anthropic's hosted Managed Agents service are separate deployment choices.
Read the documentationMicrosoft Agent Framework
Agents, explicit workflows and an opinionated harness in one toolkit.
Best fit: Enterprise teams connecting existing services, multiple model providers and managed or self-hosted agent applications.
- Graph and functional workflows
- Middleware and model-provider integrations
- Harness planning, memory and observability
The trade-off: Feature maturity differs by language and integration. Check the lifecycle of each package instead of assuming the whole ecosystem has one release status.
Read the documentationAgent Development Kit (ADK)
Build, evaluate and deploy agents with structured workflows.
Best fit: Teams wanting agent development tools, graph-based workflows and Google ecosystem integrations across several programming languages.
- Structured graph workflows
- Tools, sessions and evaluation
- Python, TypeScript, Go, Java and Kotlin
The trade-off: Check capabilities in the specific language package you choose. A framework supplies building blocks; it does not supply your process controls or success criteria.
Read the documentationCapabilities and release status can differ by language and package. These profiles were checked on 20 September 2026. Test a harness on your tasks, with your data and permission boundaries.
AGENT INTELLIGENCE
View all stories Inside the agentic shift.
Does every specialist really need its own AI agent?
A Microsoft engineering example replaces some specialist agent loops with distributed skills and MCP tools, showing how teams can preserve domain services while simplifying where reasoning happens across a workflow.
OpenAI brings the Codex harness to a managed Agents API.
OpenAI launched its Agents API in public beta, combining a managed Codex harness with configurable execution environments, context management and subagents. Teams still own the domain workflow and its outcomes.
A better handoff: deciding what a subagent should remember.
LangChain describes isolated and forked context modes in Deep Agents. Workers can inherit earlier investigation, while reviewers can begin independently. The choice changes how teams manage repeated work and bias.