AI Agents

You can put a model in a loop with tools without pretending it is a pipeline.

1. 1A model is not an agentYou can apply this in a live incident: A model predicts tokens.12m30xp
2. 2Binding a toolYou can apply this in a live incident: A tool must be bound to the model as a callable schema.12m30xp
3. 3Tool schemasYou can apply this in a live incident: JSON Schema describes the arguments the model may send.12m30xp
4. 4The agent loopYou can apply this in a live incident: Observe, decide, act, observe again.12m30xp
5. 5Scratchpads and hidden reasoningYou can apply this in a live incident: Intermediate reasoning is not the user-facing answer.12m30xp
6. 6Short-term versus long-term memoryYou can apply this in a live incident: The context window is short-term memory.12m30xp
7. 7RAG is retrieval, not actionYou can apply this in a live incident: RAG retrieves text into the prompt.12m30xp
8. 8Evaluating agentsYou can apply this in a live incident: Happy-path demos are not evaluations.12m30xp
9. 9GuardrailsYou can apply this in a live incident: Allow-list tools, not just prompt against misuse.12m30xp
10. 10Human in the loopYou can apply this in a live incident: Irreversible side effects need a confirmation gate.12m30xp

You can run agents without treating non-determinism as a personality.

1. 1Retries and idempotencyYou can apply this in a live incident: Retries need idempotency keys on the tool side.12m30xp
2. 2TracingYou can apply this in a live incident: A trace is the sequence of model calls, tool calls and observations.12m30xp
3. 3Cost and latency budgetsYou can apply this in a live incident: Each loop iteration is a model call.12m30xp
4. 4Fallback to a human or a formYou can apply this in a live incident: A tool error is a cue to stop, not to improvise.12m30xp
5. 5When to split agentsYou can apply this in a live incident: Tool overload confuses selection.12m30xp

You can choose ReAct, tools and structured output without treating the model as a database.

1. 1ReActYou can apply this in a live incident: ReAct interleaves reasoning traces with tool actions.12m30xp
2. 2Function callingYou can apply this in a live incident: Function calling is a native model API feature with a tools list.12m30xp
3. 3Structured outputYou can apply this in a live incident: Structured output constrains the model to a schema.12m30xp
4. 4Output parsersYou can apply this in a live incident: Parsers extract a typed object from model text or from native structured output.12m30xp
5. 5Planning versus reactive agentsYou can apply this in a live incident: A planner-worker split writes a plan then executes.12m30xp
6. 6System prompt as policyYou can apply this in a live incident: The system prompt states policy for the model.12m30xp
7. 7MCP at concept levelYou can apply this in a live incident: Model Context Protocol is a standard way to expose tools and resources to a host.12m30xp
8. 8Streaming tokensYou can apply this in a live incident: Streaming emits tokens before the final message exists.12m30xp
9. 9Deterministic glue versus LLMYou can apply this in a live incident: Arithmetic, routing tables and schema mapping belong in code.12m30xp
10. 10Citations versus toolsYou can apply this in a live incident: A citation should point at retrieved or tool-returned documents.12m30xp
11. 11Parallel tool callsYou can apply this in a live incident: Some models emit multiple tool calls in one turn.12m30xp
12. 12Tool choice settingsYou can apply this in a live incident: APIs can force a tool, forbid tools, or leave auto.12m30xp

You can separate session, user and vector memory without poisoning the next turn.

1. 1Session versus user memoryYou can apply this in a live incident: Session memory lives for a conversation id.12m30xp
2. 2Vector storesYou can apply this in a live incident: A vector store retrieves by embedding similarity.12m30xp
3. 3ChunkingYou can apply this in a live incident: Chunk size and overlap are retrieval hyperparameters.12m30xp
4. 4Embeddings and hybrid searchYou can apply this in a live incident: Dense embeddings capture paraphrase; sparse or keyword capture exact tokens.12m30xp
5. 5RerankingYou can apply this in a live incident: A reranker scores retrieved chunks with a more expensive model.12m30xp
6. 6Conversation compactionYou can apply this in a live incident: Long threads must be truncated, summarised or stored externally.12m30xp
7. 7Memory poisoningYou can apply this in a live incident: Writable memory is an attack surface.12m30xp
8. 8Hallucination under retrievalYou can apply this in a live incident: The model can ignore retrieved text.12m30xp
9. 9Prompt injectionYou can apply this in a live incident: Untrusted retrieved or tool text can override instructions.12m30xp
10. 10Sandboxing toolsYou can apply this in a live incident: Tool runtimes should be sandboxed: network, FS, secrets, time.12m30xp
11. 11Tool result sizeYou can apply this in a live incident: Tool results consume context tokens.12m30xp
12. 12Timeout versus successYou can apply this in a live incident: A timeout is unknown, not failure.12m30xp

You can measure traces, cost and gates without demo-driven shipping.

1. 1Eval harnessYou can apply this in a live incident: An eval harness runs fixed tasks and scores traces.12m30xp
2. 2Traces and spansYou can apply this in a live incident: A trace is one request; spans are timed steps inside it.12m30xp
3. 3Cost per stepYou can apply this in a live incident: Each model call has token cost.12m30xp
4. 4Human approval gatesYou can apply this in a live incident: A gate is a human or independent policy service, not another model sample.12m30xp
5. 5Multi-agent routerYou can apply this in a live incident: A router classifies intent to a specialist agent with fewer tools.12m30xp
6. 6LLM-as-judge pitfallsYou can apply this in a live incident: Judge models need a rubric aligned to the failure mode.12m30xp
7. 7Prompt versioningYou can apply this in a live incident: Prompts are code; they need versions and rollback.12m30xp
8. 8Red-teaming agentsYou can apply this in a live incident: Red teams try injection, exfil, and unintended tool use.12m30xp
9. 9Rate limits and circuit breakersYou can apply this in a live incident: Backoff and jitter belong in the runtime.12m30xp
10. 10Validating tool argumentsYou can apply this in a live incident: JSON Schema required fields are not enough for business rules.12m30xp
11. 11Evaluating refusalsYou can apply this in a live incident: Refusal cases belong in the golden set.12m30xp
12. 12Canarying agent changesYou can apply this in a live incident: Canary a percent of traffic or of tenants.12m30xp

You can operate retrieval, cache and handoff without silent policy drift.

1. 1Permissions in retrievalYou can apply this in a live incident: Retrieval must filter by the caller's permissions.12m30xp
2. 2Prompt and retrieval cachesYou can apply this in a live incident: Caches must include corpus version in the key.12m30xp
3. 3Handoff state to humansYou can apply this in a live incident: Handoff should include trace id, last tool results and user goal.12m30xp
4. 4PII in tracesYou can apply this in a live incident: Traces often contain user data and tool payloads.12m30xp
5. 5Model routingYou can apply this in a live incident: Route cheap models to cheap tasks.12m30xp
6. 6SLOs for agentsYou can apply this in a live incident: SLOs should include groundedness, tool-error rate and latency.12m30xp
7. 7Few-shot trapsYou can apply this in a live incident: Examples steer distribution strongly.12m30xp

You can keep an agent from becoming an unowned process.

1. 1Allow-listed toolsYou can apply this in a live incident: Each bound tool is a capability.12m30xp
2. 2Argument validationYou can apply this in a live incident: Validate arguments in the tool, not in the prompt.12m30xp
3. 3Multi-user isolationYou can apply this in a live incident: Memory keys must include the user or tenancy.12m30xp

You can bound loops, schemas and retrieval freshness in production.

1. 1Max iterationsYou can apply this in a live incident: A max-step budget is a hard stop.12m30xp
2. 2Strict schema modeYou can apply this in a live incident: Strict schema modes reject extra fields.12m30xp
3. 3Embedding model versionYou can apply this in a live incident: Query and index embeddings must be the same model and dimension.12m30xp
4. 4Metadata filtersYou can apply this in a live incident: Metadata filters are exact constraints before or after ANN.12m30xp
5. 5Agentic RAGYou can apply this in a live incident: Agentic RAG lets the model issue further searches.12m30xp
6. 6Self-critique and reflectionYou can apply this in a live incident: Reflection is another model call with a rubric.12m30xp
7. 7Temperature with toolsYou can apply this in a live incident: High temperature increases format errors on tool calls.12m30xp
8. 8Abstain behaviourYou can apply this in a live incident: Abstain must be trained with evals and sometimes forced when retrieval is empty.12m30xp
9. 9Per-tool timeoutsYou can apply this in a live incident: Each tool needs a timeout shorter than the user SLO.12m30xp
10. 10Session isolationYou can apply this in a live incident: Conversation state must be keyed per session.12m30xp
11. 11Tool-selection evalsYou can apply this in a live incident: Score which tool was chosen, not only the prose.12m30xp
12. 12Cost alertsYou can apply this in a live incident: Provider budgets and local step caps are both required.12m30xp