Agents
8 current practices, and 4 retired.
StandardAgentsAG-001Define an agent as a model using tools in a loopThe core primitive is a model that autonomously calls tools until the task is done. Everything else is workflow.
Reviewed Aug 2026StandardAgentsAG-002Design tools for agents, not for developersNamespace related tools, write descriptions for a non-deterministic reader, and return token-efficient, high-signal results.
Reviewed Aug 2026AgentsAG-003Use an orchestrator with parallel subagents for open-ended researchA lead agent plans and spawns parallel subagents, each with its own context window, then synthesises their findings.Reviewed Aug 2026AgentsAG-004Let the agent write code that calls toolsRather than issuing many separate tool calls into context, have the agent write code that calls tools and consumes the intermediate results.
Reviewed Aug 2026StandardAgentsAG-005Use MCP as the standard integration layerImplement the Model Context Protocol once and inherit an ecosystem of tool and data integrations.Reviewed Aug 2026AgentsAG-006Make human approval a tool the agent callsModel human input and approval as a tool call rather than a hard-coded stop.
Reviewed Aug 2026AgentsAG-007Own your prompts, context and control flowKeep explicit control of what goes into the model and how control flows, instead of ceding it to framework abstractions.Reviewed Aug 2026StandardAgentsAG-008Keep a durable session log and a disposable harnessStore the canonical event log durably and treat the agent loop and sandbox as cheap and recreatable.
Reviewed Aug 2026
Reviewed Aug 2026StandardAgentsAG-002Design tools for agents, not for developersNamespace related tools, write descriptions for a non-deterministic reader, and return token-efficient, high-signal results.Reviewed Aug 2026AgentsAG-003Use an orchestrator with parallel subagents for open-ended researchA lead agent plans and spawns parallel subagents, each with its own context window, then synthesises their findings.Reviewed Aug 2026AgentsAG-004Let the agent write code that calls toolsRather than issuing many separate tool calls into context, have the agent write code that calls tools and consumes the intermediate results.Reviewed Aug 2026StandardAgentsAG-005Use MCP as the standard integration layerImplement the Model Context Protocol once and inherit an ecosystem of tool and data integrations.Reviewed Aug 2026AgentsAG-006Make human approval a tool the agent callsModel human input and approval as a tool call rather than a hard-coded stop.Reviewed Aug 2026AgentsAG-007Own your prompts, context and control flowKeep explicit control of what goes into the model and how control flows, instead of ceding it to framework abstractions.Reviewed Aug 2026StandardAgentsAG-008Keep a durable session log and a disposable harnessStore the canonical event log durably and treat the agent loop and sandbox as cheap and recreatable.Reviewed Aug 2026Retired in this pillar
RetiredAgentsRE-004Few-shot examples to force a JSON tool callFew-shot prompting to get a model to emit correctly-shaped tool-call JSON.
Retired Jun 2023RetiredAgentsRE-006Add and remove tools mid-task to keep the action space smallLoading tools on demand, RAG-style, so the agent sees fewer options at once.
Retired Jul 2025RetiredAgentsRE-010Reach for a heavy agent framework firstStarting a production agent by adopting a full orchestration framework.Retired Dec 2024RetiredAgentsRE-012Hand-wire many individual tool calls into one contextManually orchestrating long sequences of tool calls through the prompt.Retired 2025
Retired Jun 2023RetiredAgentsRE-006Add and remove tools mid-task to keep the action space smallLoading tools on demand, RAG-style, so the agent sees fewer options at once.Retired Jul 2025RetiredAgentsRE-010Reach for a heavy agent framework firstStarting a production agent by adopting a full orchestration framework.Retired Dec 2024RetiredAgentsRE-012Hand-wire many individual tool calls into one contextManually orchestrating long sequences of tool calls through the prompt.Retired 2025