Choose and operate a model
Which model family and operating pattern fits my workload?
Compare current model families by their documented controls and learning curves, then evaluate them on your own representative tasks.
AgentAtlas defaultReviewed 2026-07-26
Start with the smallest production-ready agent stack.
There is no universal best stack. This is the low-regret default: give the agent durable context, add only the capabilities the job needs, verify outcomes, and keep permissions narrow.
Default path
Build a production agent
Start with durable repository context, add reusable workflows and tools only where they earn their complexity, then require verification and least-privilege permissions.
- 1Production Agent Starter KitCreate the reviewed instructions, specification, plan, task, runtime, safety, and operations files the project actually needs.
- 2Repository instructionsGive every compatible agent the same project map, commands, conventions, and boundaries.
- 3SkillsPackage repeatable workflows after the repository baseline is stable.
- 4MCP and toolsConnect only the external systems the workflow actually needs.
- 5Evals and verificationDefine success checks before expanding autonomy or deployment scope.
- 6Permissions and safetyGrant the narrowest access that can complete the verified workflow.
Route by goal
Skip directly to the decision you need.
Connect tools and data
Should this capability be a Skill, plugin, or MCP integration?
Keep repeatable procedure in a Skill, use the host's plugin packaging when distribution requires it, and use MCP for runtime access to external systems.
Coordinate multiple agents
Do I need subagents, automations, MCP, or A2A?
Use subagents for bounded delegation inside one run, automations for scheduled or event-driven work, and A2A when independently operated agents need to interoperate.
Manage instructions and memory
Where should durable context, learned memory, and event logic live?
Keep stable project rules in repository instructions, store learned state in an explicit memory layer, and reserve hooks for deterministic lifecycle behavior.
Automate repeatable work
How should I turn a successful agent run into a repeatable system?
Extract the procedure into a Skill, add deterministic hooks where required, then move it into an automation only when its trigger, retries, ownership, and observability are explicit.
Ship with evidence and guardrails
What is the minimum safety and quality layer for production?
Use task-representative evals, explicit completion checks, narrow permissions, and human approval at consequential boundaries.
Enable agent commerce
How should an agent discover offers and authorize transactions?
Separate offer discovery, checkout coordination, user authorization, payment execution, and receipts into auditable protocol boundaries.
Recommendations you can inspect.
- Rules-based and non-personalized: the same goal produces the same path.
- Primary documentation is linked through every guide's evidence ledger.
- Review dates and maximum review intervals are exposed to people and machines.
- Commercial partners cannot purchase rankings, conclusions, or recommendation positions.
For agents, bots, and AI search
Call the recommendation endpoint without a goal for the default path, or pass a listed goal. Responses include canonical URLs, reasons, sponsorship disclosures, freshness, and every available route.