Cheatsheet

AI agent design cheatsheet (2026 patterns that ship)

Q: Best AI agent framework in 2026?

LangGraph for graph-based; OpenAI Agents SDK if OpenAI-native; Claude Agent SDK if Anthropic-native; CrewAI for multi-agent. All credible.

Q: How many tools is too many for one agent?

Past ~30 tools per agent, mis-routing climbs. Split into specialised sub-agents instead.

Q: Should I use a planner-executor or ReAct?

ReAct for ≤5 steps; planner-executor for longer horizons.

Production-tested AI agent design patterns in 2026: planner-executor split, tool design, context distillation, budget caps, eval discipline, and the antipatterns that bite.

Architecture patterns

Pick a pattern based on horizon and branching.

Item	Description	Example
`ReAct loop`	Thought → action → observation, repeat. Good for short horizons (≤5 steps).	`Quick research, single-tool retrieval`
`Planner-executor`	Planner decomposes goal into subtasks; executor runs each with its own budget. Default for >7 steps.	`Multi-file code refactor`
`Graph (state machine)`	Explicit nodes + transitions; LangGraph default. Easiest to debug.	`Workflow with branching, retries`
`Critic-actor`	Critic scores output; actor revises until critic approves. Quality over cost.	`Code generation with iterative fix loops`

Tool design rules

Item	Description	Example
`Name by intent`	search_company_news, not http_get_news_api. The model uses names as docs.
`Description = documentation`	1-2 sentences per tool. Include when NOT to use it.
`Tight schemas`	JSON Schema with descriptive field names. Required vs optional matters.
`Readable errors`	"Invalid ticker 'XYZ' — try 4-letter ticker like 'AAPL'" beats "400 Bad Request".
`Small results`	Return summaries, not 10K-token JSON. The model has to fit them in next step's context.
`Tool count <30`	Past that, mis-routing climbs. Split into sub-agents.

Context engineering

Item	Description	Example
`Distill the loop`	Every N steps, rewrite history as a tight summary. Drop verbatim tool outputs after extracting what matters.
`System prompt for constants`	Tools, role, format — anything stable. Maximises prompt caching.
`User turn for variables`	Task, latest tool result, next step.
`Beware lost-in-middle`	Put critical instructions at head + tail of the prompt.

Budgets + guardrails (mandatory)

Item	Description	Example
`Max-step cap`	Hard ceiling on loop iterations (e.g. 20). Set at framework level.
`Token budget`	Total tokens across the loop. Kill if exceeded.
`Wall-clock budget`	Total time. Important for user-facing agents.
`No-progress detector`	If plan / summary hasn't changed in N steps, stop and escalate.
`Refusal whitelist`	Explicit list of actions the agent must never take.

Routing

Item	Description	Example
`Cheap router model`	GPT-4o-mini / Claude Haiku picks which tool to call. Fast, near-free.
`Strong executor model`	Claude 4.6 / GPT-4o runs the actual step.
`Reasoning model selectively`	Only for planning + hardest subtasks. Save the budget.

Eval discipline

Item	Description	Example
`End-to-end success`	Score on 50-200 golden tasks. The metric that matters.
`Step-level correctness`	Was each tool right? Were args right? For debugging.
`Failure mode taxonomy`	Wrong tool, wrong args, no progress, hallucinated result, timeout. Count each.
`Regression alarm`	Run evals on every prompt / tool change. Block deploy on N% drop.

FAQ

Best AI agent framework in 2026?