AI Agents

ReAct (Reason + Act)

• Interleaves reasoning traces (Thought) with tool calls (Action) and results (Observation)
• Grounding reasoning in tool outputs reduces hallucination compared to pure chain-of-thought
• Stopping condition: agent generates a final answer or reaches a maximum step count
• ReAct is the baseline pattern behind most production agents

Plan-and-Execute

• Planner model generates a high-level task list; executor model carries out individual steps
• Separating planning from execution allows specialized prompts and smaller executor models
• Dynamic replanning: executor can request revised plan if an earlier step fails
• Suitable for long-horizon tasks where the full plan can be specified upfront

Task Decomposition

• Break large tasks into atomic subtasks the model can reliably complete
• Decomposition reduces per-step ambiguity and makes evals easier
• Tree-of-Thought explores multiple decompositions and selects the best branch
• Subtask granularity should match tool capabilities — too coarse → tool misuse, too fine → overhead

Tool Design Principles

• Each tool should do one thing well — single responsibility reduces misuse
• Return structured data from tools so the model can parse results reliably
• Include explicit error messages and suggested recovery actions in tool outputs
• Idempotent tools are safer — agents may call them multiple times on retries

Tool Selection Accuracy

• Accuracy degrades as tool count increases beyond ~10–15 tools in one prompt
• Categorize tools and dynamically load only relevant ones per query
• Fine-tuning the model on tool-use examples significantly improves selection accuracy
• Measure tool selection precision and recall on a labeled evaluation set

Agent Memory Types

• In-context: full conversation history in the prompt; limited by context window
• External (episodic): retrieve past interactions from a vector store
• Semantic: distilled summaries or knowledge stored in structured form
• Procedural: learned behaviors encoded in fine-tuned weights or system prompt instructions

Context Window Management

• Summarize or truncate older turns when context fills up
• Keep the most recent N turns and a running summary of earlier history
• Moving summary window: update the summary incrementally as new turns arrive
• Retrieving relevant past episodes on demand outperforms naive full-history truncation

Agentic Loop Control

• Every agent loop must have a maximum step count / timeout to prevent infinite loops
• Termination conditions: task complete, explicit STOP tool call, or error threshold exceeded
• Checkpointing intermediate state lets you resume or replay from a known good point
• Human-in-the-loop gates can interrupt the loop for high-stakes decisions

Determinism and Reproducibility

• Set temperature=0 for agents where reproducibility matters more than diversity
• Log every tool call and model response for debugging and replay
• Seed randomness in any sampling steps for reproducible test runs

Error Handling Patterns

• Return typed errors from tools with a 'how to fix this' field in the schema
• Retry with exponential backoff on transient tool failures (network, rate limits)
• If retries fail, escalate to a different tool or ask for human clarification
• Track consecutive failures; abort and alert after N failed attempts

Sandboxing and Side-Effect Management

• Run code execution tools in isolated environments (Docker, Firecracker microVMs)
• Limit tool permissions to least privilege: read-only file access unless write is needed
• Dry-run mode: simulate irreversible actions before committing
• Audit log every write operation for compliance and rollback

Agent Evaluation Challenges

• Final-answer accuracy alone misses planning and tool-use quality
• Trajectory evaluation: score each step (correct tool?, correct args?, correct handling of result?)
• Task completion rate and steps-to-completion are key throughput metrics
• Regression suites must cover diverse task types to catch capability drift

Offline vs Online Agent Evals

• Offline: replay recorded tool responses so evals run fast and cheaply
• Online: run against real tools; more realistic but slower and costlier
• Shadow mode: run new agent version in parallel with production; compare outcomes
• LLM-as-judge can score trajectory quality at scale when human annotation is too expensive