Agent architectures that real teams ship

The taxonomy most production teams reason with comes from Anthropic's "Building effective agents", published 19 December 2024. It draws a hard line between workflows, described as systems where LLMs and tools are orchestrated through predefined code paths, and agents, systems where LLMs dynamically direct their own processes and tool usage. That line is the first architectural decision on any project, and it is usually the one that decides cost and debuggability.

The research primitive underneath all of it is ReAct (arXiv 2210.03629, October 2022), which interleaved reasoning traces with tool actions. With one or two in-context examples it outperformed imitation and reinforcement learning baselines by 34 and 10 absolute success-rate points on ALFWorld and WebShop respectively. Most agent runtimes shipping today are recognisable descendants of that loop.

The practical value of the workflow list is that all five patterns keep control flow in code rather than in the model. They have predictable token counts, deterministic control flow and stack traces. Reaching for a dynamic agent when a router or an evaluator-optimiser would do is an avoidable cost, and the token multiples below show why.

Anthropic's Research feature, described on 13 June 2025, is a concrete orchestrator-workers deployment: a Claude Opus 4 lead agent spawning Claude Sonnet 4 subagents in parallel, with a separate citation agent. It outperformed single-agent Opus 4 by 90.2% on Anthropic's internal research eval. That is a large and well-documented gain for a class of task that genuinely parallelises.

The cost side is the part most teams underweight. Anthropic reports that agents use roughly 4x the tokens of a chat interaction, and multi-agent systems roughly 15x. On BrowseComp, token usage alone explained 80% of performance variance, which means much of what looks like architectural cleverness is spending.

Read those two facts together and the rule falls out. Multi-agent is defensible where the subtasks are genuinely independent and each subagent returns a distilled summary of roughly 1,000 to 2,000 tokens. Everywhere else, a single thread is cheaper and easier to explain when it goes wrong.

The day before Anthropic's post, on 12 June 2025, Cognition published "Don't Build Multi-Agents". It argues for a single-threaded linear agent with a dedicated context-compression model instead of parallel subagents. Its two principles name the failure surface precisely: share full agent traces rather than individual messages, and treat every action as carrying an implicit decision that can conflict with another agent's.

The empirical version came from Berkeley. "Why Do Multi-Agent LLM Systems Fail?" (arXiv 2503.13657, submitted 17 March 2025, v3 26 October 2025) built MAST, a taxonomy of 14 distinct failure modes across three categories: system design, inter-agent misalignment, and task verification. It was derived from 150 expert-annotated traces at inter-annotator agreement kappa=0.88, then extended to over 1,600 traces across 7 multi-agent frameworks.

Neither position is a rebuttal of the other. They describe the same trade: parallelism buys coverage and costs coherence. The useful move is to treat MAST as a pre-mortem checklist rather than as an argument, and to design the verification step before the delegation step.

Anthropic's "Effective context engineering for AI agents" (29 September 2025) reframed the discipline as curating and maintaining the optimal set of tokens during inference, rather than writing better prompts. It named the mechanisms now standard in production runtimes: just-in-time retrieval via lightweight identifiers, compaction at the context limit, structured note-taking, and sub-agent context isolation.

The premise behind it was measured. Chroma's "Context Rot" study of 14 July 2025 tested 18 frontier models and found performance degrades as input grows, even on trivial tasks, well before the nominal context window is full. Degradation was sensitive to needle-question similarity, to distractors and to haystack structure. A long window is a budget, not a guarantee.

The 2026 term of art is harness engineering. Mitchell Hashimoto named the stage on 5 February 2026, defining it as taking the time, whenever an agent makes a mistake, to engineer a solution so that the agent never makes that mistake again, implemented through AGENTS.md-style implicit prompting and purpose-built verification scripts. Lilian Weng formalised it on 4 July 2026 as the system that orchestrates execution and decides how the model thinks and plans, calls tools and acts, perceives and manages context, stores artifacts, and evaluates results.

It is now a product surface in preview. Microsoft's Agent Harness, announced at Build on 3 June 2026 and described as the layer where model reasoning meets real execution, bundles shell and filesystem access, plan-versus-execute modes, automatic compaction, session-scoped file memory, skill discovery and child-agent delegation. The same capabilities that teams hand-rolled through 2025 are becoming platform features.

A closely related shift is to let the model write code against a tool API rather than emit one tool call at a time. Anthropic's 4 November 2025 post showed an MCP workflow dropping from 150,000 tokens to 2,000, a 98.7% reduction, by exposing servers as an on-disk code API the model reads on demand. The saving comes from not loading every tool definition and every intermediate result into context.

Microsoft measured the same effect independently and published it at Build on 3 June 2026. CodeAct-style orchestration cut a multi-step workload from 27.81 seconds to 13.23 seconds, a 52.4% reduction in execution time, and from 6,890 to 2,489 tokens, a 63.9% reduction, against traditional tool-call orchestration.

Two vendors measuring the same direction on different stacks is a stronger signal than either result alone. It also changes the security question: a code-executing agent needs a sandbox and an explicit permission model, not just a tool allowlist.

LangChain and LangGraph both reached v1.0 on 22 October 2025 at 90 million monthly downloads, naming Uber, JP Morgan, BlackRock, Cisco, LinkedIn and Klarna as production users. LangGraph's specific contribution is durable state, checkpointing and first-class human-in-the-loop, which are the parts teams otherwise rebuild badly.

Microsoft collapsed AutoGen and Semantic Kernel into Agent Framework 1.0 on 3 April 2026, an MIT-licensed SDK for .NET and Python with both MCP and A2A support, shipping sequential, concurrent, handoff, group-chat and Magentic-One orchestration. Google's Agent Development Kit reached 1.0.0 for Java and Go, with Python at 2.0 beta, by 21 May 2026. CrewAI, per its investor Insight Partners on 10 December 2025, reports 1.4 billion agentic automations powered, roughly 450 million agents running monthly, and use across more than 60% of the Fortune 500.

OpenAI's trajectory is the caution. On 3 June 2026 it announced that the visual Agent Builder and Evals are deprecated, both scheduled to shut down on 30 November 2026, with migration to the Agents SDK or ChatGPT Workspace Agents. Visual builders are a convenient on-ramp and a poor place to keep a system you depend on.

Google donated Agent2Agent to the Linux Foundation on 23 June 2025 at Open Source Summit North America, with AWS, Cisco, Microsoft, Salesforce, SAP and ServiceNow among the launch supporters. A2A reached v1.0, its first production-ready stable release, in March 2026, adding signed Agent Cards for cryptographic agent identity verification.

Anthropic donated the Model Context Protocol to the newly formed Agentic AI Foundation on 9 December 2025, alongside Block's goose and OpenAI's AGENTS.md. At donation, MCP reported over 97 million monthly SDK downloads and roughly 10,000 active servers. The foundation's platinum tier includes AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI.

Observability is the least settled layer. The OpenTelemetry GenAI semantic conventions define an invoke_agent span with child chat and execute_tool spans, plus attributes such as gen_ai.request.model and token-usage counters, but OpenTelemetry describes them as still under active development rather than stable. That is enough to instrument against today, provided you expect attribute names to move.

Klarna's assistant is the most-cited case and the most instructive. On 27 February 2024 Klarna reported 2.3 million conversations in the first month, two-thirds of its customer service chats, doing work equivalent to 700 full-time agents, with resolution time falling from 11 minutes to under 2 and a projected $40 million profit improvement for 2024. LangChain's case study of 12 February 2025 attributes the architecture to LangGraph multi-agent routing by task type, and reports an 80% cut in average query resolution time across 2.5 million conversations after nine months. Press coverage in 2026, including Forbes on 16 July 2026, reported that Klarna later reintroduced human agents for complex cases after cutting too aggressively.

Two other deployments give comparable shapes. A Salesforce customer story dated 4 December 2024 reports that Agentforce on the Salesforce Help site resolves roughly 76% of customer inquiries without a human and escalates only 5% to a support engineer, across more than 1.7 million Agentforce conversations. Intercom reported on 12 March 2026 a 76% average Fin resolution rate across more than 12,000 teams, while moving billing from resolutions to outcomes, a pricing change that itself signals how often an agent completes part of a task and hands off. Databricks reported over 100,000 agents built and more than one quadrillion agent tokens processed per year as of 16 June 2026.

Now the discount. Gartner predicted on 25 June 2025 that over 40% of agentic AI projects will be cancelled by end-2027, citing escalating costs, unclear business value and inadequate risk controls. LangChain's State of Agent Engineering, based on 1,340 respondents surveyed from 18 November to 2 December 2025, found 57.3% with agents in production, rising to 67% at organisations with 10,000 or more employees, but with quality rather than cost named as the top blocker at 33%. Adoption is real; the failure rate is also real, and it is a quality problem before it is a budget problem.

The decision sequence that follows from the evidence is short. Classify the work against the five workflow patterns first, and only reach for a dynamic agent when the path genuinely cannot be written in advance. Use multi-agent where the work parallelises and each subagent can return a distilled summary of roughly 1,000 to 2,000 tokens; use a single thread otherwise, as Cognition argued on 12 June 2025.

Then build the harness rather than the prompt. The Claude Agent SDK structures agents around a four-phase loop, gather context, take action, verify work, iterate, and MAST's task-verification category is the reason to treat the verify phase as a subsystem rather than a final prompt. Anthropic's context engineering mechanisms, just-in-time retrieval, compaction, structured notes and sub-agent isolation, are the concrete implementation of the gather phase.

Finally, keep the seams open. MCP for tool supply and A2A for cross-agent delegation are both foundation-governed now, code-first SDKs survive product retirements that visual builders do not, and OpenTelemetry GenAI spans give you a portable trace even while the conventions are still moving. None of this makes an agent work. It makes the parts you get wrong cheaper to replace.