Mechanics Behind Autonomous Systems

Resources

• Building Effective Agents — Anthropic distinguishes predefined workflows from agents that choose their own steps, then explains stopping conditions and human intervention. A useful starting point for deciding how much independence a task warrants rather than treating autonomy as an on/off switch. https://www.anthropic.com/engineering/building-effective-agents

• Guidelines for Human–AI Interaction — Microsoft Research’s original guidelines address capability disclosure, correction, dismissal, and user control. Use them to connect supervision levels and trust calibration to what the operator can actually understand and interrupt. https://www.microsoft.com/en-us/research/publication/guidelines-for-human-ai-interaction/

• Effective Harnesses for Long-Running Agents — Anthropic describes initialization, progress artifacts, incremental work, and clean handoffs between context windows. This makes continuity concrete: persistent files and verifiable checkpoints matter more than simply keeping a model loop alive. https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents

• Temporal Workflow Execution — The official explanation of durable workflow execution introduces event history and recovery across worker failures. Use it to distinguish durable orchestration state from an agent’s conversational memory and from the external effects its tools create. https://docs.temporal.io/workflow-execution

• Temporal Task Queues — Explains how workers receive workflow and activity tasks through queues. Provides the runtime foundation for separating incoming work, available execution capacity, and an autonomous agent’s decision loop. https://docs.temporal.io/task-queue

• Temporal Schedules — Covers recurring workflow activation and policies for overlapping or missed runs. Useful for teaching proactive behavior as explicit scheduled work with operational rules, not an endlessly polling prompt. https://docs.temporal.io/schedule

• Temporal Activity Execution — Explains activity execution, cancellation, and asynchronous completion, including the role of heartbeats in receiving cancellation. Use it to show why a long-running job needs a cooperative stop protocol and cleanup behavior. https://docs.temporal.io/activity-execution

• LangGraph Persistence — Documents threads, checkpoints, stores, and recovery from saved execution state. Helps distinguish per-run state from shared cross-run memory and identify what must survive a process restart. https://docs.langchain.com/oss/python/langgraph/persistence

• LangGraph Interrupts — Shows how a workflow pauses for human input and resumes using persisted state. Particularly useful for approval gates and escalation, including the warning that resumed nodes can re-execute code before the interruption. https://docs.langchain.com/oss/python/langgraph/interrupts

• Running Agents with the OpenAI Agents SDK — Documents the execution loop, run configuration, continuation, and turn limits. Use it to separate a bounded agent run from the surrounding supervisor that must enforce deadlines, monetary budgets, and escalation rules. https://openai.github.io/openai-agents-python/running_agents/

• OpenAI Agents SDK Tracing — Covers traces and spans for model calls, tools, guardrails, and handoffs, with controls for sensitive data. Provides an inspection surface for progress and failures; pair traces with usage accounting rather than assuming observability itself enforces cost limits. https://openai.github.io/openai-agents-python/tracing/

• Timeouts, Retries, and Backoff with Jitter — The Amazon Builders’ Library explains how retries can amplify overload and why timeouts, bounded retries, and jitter belong together. Apply these mechanisms to model and tool failures before granting a system unattended retry authority. https://aws.amazon.com/builders-library/timeouts-retries-and-backoff-with-jitter/

• Compensating Transactions — Microsoft’s architecture pattern explains how to recover from partially completed operations when a database-style rollback is unavailable. Useful for distinguishing checkpoint restoration from repairing external side effects such as notifications, reservations, or account changes. https://learn.microsoft.com/en-us/azure/architecture/patterns/compensating-transaction

• OWASP Excessive Agency — Identifies excessive functionality, permissions, and autonomy as distinct sources of risk. Supports concrete control boundaries: narrow tools, least privilege, independent authorization, human approval, and limits on repeated actions. https://genai.owasp.org/llmrisk/llm062025-excessive-agency/