Agent Engineering

Every agent in 2026 is a variant of the ReAct loop from 2022 — Claude Code, Cursor, Devin, Operator included. Reasoning tokens interleave with tool calls and observations until a stop condition fires. Learn this loop cold before touching any framework.

ReAct interleaves thought and action in one stream. ReWOO separates them: one big plan up front, then execute. 5x fewer tokens, +4% accuracy on HotpotQA, and you can distill the planner into a 7B model. Plan-and-Execute generalized it; Plan-and-Act scaled it to web navigation.

Gradient-based RL needs thousands of trials and a GPU cluster to fix a failure mode. Reflexion (Shinn et al., NeurIPS 2023) does it in natural language: after each failed trial, the agent writes a reflection, stores it in episodic memory, and conditions the next trial on that memory. This is the pattern behind Letta's sleep-time compute, Claude Code's CLAUDE.md learnings, and pro-workflow's learn-rule.

A single chain-of-thought trajectory has no room to backtrack. ToT (Yao et al., 2023) turns reasoning into a tree with self-evaluation on each node. LATS (Zhou et al., 2024) unifies ToT with ReAct and Reflexion under Monte Carlo Tree Search. Game of 24 goes from 4% (CoT) to 74% (ToT); LATS hits 92.7% pass@1 on HumanEval.

Self-Refine (Madaan et al., 2023) uses one LLM in three roles — generate, feedback, refine — in a loop. Average gain: +20 absolute on 7 tasks. CRITIC (Gou et al., 2023) hardens the feedback step by routing verification through external tools. In 2026 this pattern ships in every framework as "evaluator-optimizer" (Anthropic) or a guardrail loop (OpenAI Agents SDK).

Toolformer (Schick et al., 2023) started self-supervised tool annotation. Berkeley Function Calling Leaderboard V4 (Patil et al., 2025) sets the 2026 bar: 40% agentic, 30% multi-turn, 10% live, 10% non-live, 10% hallucination. Single-turn is solved. Memory, dynamic decision-making, and long-horizon tool chains are not.

Context windows are finite. Conversations, documents, and tool traces are not. The fix is OS virtual memory restated — main context is RAM, external store is disk, the agent pages between them. MemGPT (Packer et al., 2023) named the pattern; many production memory systems build on it.

Discrete functional memory blocks the model can edit directly, and a sleep-time agent that consolidates memory asynchronously while the primary agent is idle. These two ideas are how you scale memory beyond one conversation.

Hybrid memory runs three stores in parallel — vector for semantic similarity, KV for fast fact lookup, graph for entity-relationship reasoning — with a scoring layer that fuses them on retrieval. This is a widely used production pattern for external memory; Mem0 (Chhikara et al., 2025) is one reference implementation.

Voyager (Wang et al., TMLR 2024) treats executable code as a skill. Skills are named, retrievable, composable, and refined by environment feedback. This is the reference architecture for Claude Agent SDK skills, skillkit, and the 2026 skill-library pattern.

Symbolic planning handles the cases where the plan is provably correct. Evolutionary code search handles the cases where the fitness function is machine-checkable. ChatHTN (2025) and AlphaEvolve (2025) show what each unlocks when paired with an LLM.

Schluntz and Zhang (Anthropic, Dec 2024) distinguish workflows (predefined paths) from agents (dynamic tool-use). Five workflow patterns cover most cases. Start with direct API calls. Add agents only when steps cannot be predicted.

Agent is a state machine; nodes are functions; edges are transitions; state is checkpointed after each node. Resume from any failure at the last successful checkpoint. LangGraph is the 2026 reference for this model of low-level stateful orchestration.

Agents as actors: async message exchange, event-driven handlers, fault isolation, natural concurrency. AutoGen v0.4 (Microsoft Research, Jan 2025) redesigned agent orchestration around this model; the framework is now in maintenance mode, with Microsoft Agent Framework (public preview Oct 2025) as its production successor.

Four primitives: Agent, Task, Crew, Process. Two top-level shapes: Crews (autonomous, role-based collaboration) and Flows (event-driven, deterministic). CrewAI is the 2026 reference implementation, and its docs are blunt: "for any production-ready application, start with a Flow."

OpenAI Agents SDK is the lightweight multi-agent framework built on the Responses API. Five primitives: Agent, Handoff, Guardrail, Session, Tracing. Handoffs are tools named `transfer_to_<agent>`. Guardrails trip on input or output. Tracing is on by default.

A harness you can import: built-in tools, subagents for context isolation, hooks, W3C trace propagation, session persistence. The Claude Agent SDK is the reference example — the library form of the Claude Code harness — and Claude Managed Agents is the hosted alternative for long-running async work.

A production agent runtime optimizes what prototyping frameworks ignore: instantiation cost, typed workflow surfaces, and a serving-ready backend. The 2026 pairing: Agno (Python) aims at microsecond agent instantiation and stateless FastAPI backends. Mastra ships agents, tools, workflows, unified model routing, and composite storage on the Vercel AI SDK substrate.

Three benchmarks anchor agent evaluation in 2026. SWE-bench tests code patching. GAIA tests generalist tool use. AgentBench tests multi-environment reasoning. Know their composition, their contamination story, and what they do not measure.

WebArena tests web-agent capability across four self-hosted apps. OSWorld tests desktop-agent capability across Ubuntu, Windows, macOS. At release (2023–2024) both showed a big gap between best-in-class agents and humans. The gap is narrowing; the failure modes haven't changed.

Three production computer-use models in 2026. All three are vision-based. All three treat screenshots, DOM text, and tool outputs as untrusted input. Only direct user instructions count as permission. Per-step safety services are the norm.

Voice agents are a first-class production category in 2026. Pipecat gives you a Python frame-based pipeline (VAD → STT → LLM → TTS → transport). LiveKit Agents bridges AI models to users over WebRTC. Production latency targets land at 450–600ms end-to-end for premium stacks.

OpenTelemetry's GenAI SIG (launched April 2024) defines the standard schema for agent telemetry. Span names, attributes, and content-capture rules converge across vendors so agent traces mean the same thing in Datadog, Grafana, Jaeger, and Honeycomb.

Three open-source agent observability platforms dominate 2026. Langfuse (MIT) — 6M+ installs/month, tracing + prompt management + evals + session replay. Arize Phoenix (Elastic 2.0) — deep agent-specific evals, RAG relevancy, OpenInference auto-instrumentation. Comet Opik (Apache 2.0) — automated prompt optimization, guardrails, LLM-judge hallucination detection.

Du et al. (ICML 2024, "Society of Minds") run N model instances that independently propose answers, then iteratively critique each other over R rounds to converge. Improves factuality, rule-following, reasoning. Sparse topology beats full mesh on token cost.

MASFT (Berkeley, 2025) catalogs 14 multi-agent failure modes in 3 categories. Microsoft's Taxonomy documents how existing AI failures amplify in agentic settings. Industry field data converges on five recurring modes: hallucinated actions, scope creep, cascading errors, context loss, tool misuse.

Greshake et al. (AISec 2023) established indirect prompt injection as the defining agent security problem. Attacker plants instructions in data the agent retrieves; on ingest, those instructions override the developer prompt. Treat all retrieved content as arbitrary code execution on the tool-use surface.

Four orchestration patterns recur across 2026 frameworks: supervisor-worker, swarm / peer-to-peer, hierarchical, debate. Anthropic's guidance: "It's about building the right system for your needs." Start simple; add topology only when a single agent plus five workflow patterns is insufficient.

Production agents run on six runtime shapes: request-response, streaming, durable execution, queue-based background, event-driven, and scheduled. Pick the shape before you pick the framework. Observability is load-bearing at every shape.

Anthropic's guidance: "start with simple prompts, optimize them with comprehensive evaluation, and add multi-step agentic systems only when needed." Evaluation is not the last step. It's the outer loop that drives every other choice in Phase 14.

A capable model is not enough. Reliable agents need a workbench: instructions, state, scope, feedback, verification, review, and handoff. Strip those away and even a frontier model produces work that is unsafe to ship.

The smallest useful workbench is three files: a root instructions router, a state file, and a task board. Everything else is layered on top. If a repo cannot carry these three, no model will save it.

Instructions written as prose are wishes. Instructions written as constraints are tests. The workbench turns each rule into something an agent can check at runtime and a reviewer can verify after the fact.

Chat history is volatile. The repo is durable. The workbench stores agent state in versioned files so the next session, the next agent, and the next reviewer all read from the same source of truth.

Every session that starts cold pays a tax. The agent reads the same files, retries the same probes, and rediscovers the same paths. An init script pays the tax once and writes the answers into state.

The model does not know where the work ends. A scope contract is a per-task file that says where the work begins, where it ends, and how to roll back if it spills. The contract turns "stay in scope" from a wish into a check.

Agents that do not see real command output guess. A feedback runner captures stdout, stderr, exit code, and timing into a structured record the next turn can read. Then the agent reacts to facts instead of to its own prediction of facts.

The agent does not get to mark its own work as done. A verification gate reads the scope contract, the feedback log, the rule report, and the diff, and answers a single question: is this task actually complete? If the gate says no, the task is not done, no matter what the chat says.

The agent that wrote the code cannot grade it. A reviewer is a second loop with a different system prompt, a different goal, and read-only access to everything the builder produced. The gap between builder and reviewer is where most reliability lives.

The session is going to end. The work is not. The handoff packet is the artifact that turns "the agent worked for an hour" into "the next session is productive in the first minute." Build it on purpose, not as an afterthought.

Eleven lessons of surfaces are worth nothing if they do not survive contact with a real codebase. This lesson runs the same task twice on a small sample app: prompt-only versus workbench-guided. The numbers do the arguing.

The mini-track ends with a pack you drop into any repo. Eleven lessons of surfaces compressed into a directory you can `cp -r` and have an agent working reliably the next morning. The capstone is the artifact this curriculum trades on.

A coding agent can implement a clear task quickly. It can also implement an unclear task quickly. The speed is the same. The cost is not.

A plan is not a prettier to-do list. It is a dependency graph in which every change has a reason and every terminal node has proof.

Parallel agents save wall time only when the work is independent. Otherwise they convert one clear task into a coordination problem with a faster failure rate.

A correction that lives only in chat fixes one run. A correction promoted into a test, boundary, example, or tool improves every later run.

Fast implementation increases the penalty for choosing the wrong problem. Shape the outcome first so speed points in the right direction.

Requirements are not waiting in a meeting to be collected. They are scattered across actions, workarounds, records, and disagreements.

A roadmap hides uncertainty inside features. An assumption map exposes what must be true before those features deserve to exist.

Small is useful only when it proves something important. A tiny build that cannot change the next decision is merely incomplete.

A useful specification fixes invariants and evidence while leaving reversible implementation choices open. It is a decision boundary, not a screenplay.

Measurement should answer a decision, not decorate a dashboard. Start with the goal, derive questions, then choose the smallest metrics that answer them.

These are different learning environments, not levels of polish. Choose the stage that answers the current unknown with the least unnecessary consequence.

Shipping closes one build loop and opens the learning loop. Evidence must change the system or it becomes telemetry nobody owns.