Agentic coding has stopped being a demo and started becoming an architectural choice — one with a price tag, a hardware footprint, and a vocabulary problem. The state of the art looks like this: an Orchestrator agent delegating to a 'pantheon' of specialists, each pinned to whichever model is cheapest for its role; a 512GB Mac Studio in the corner running a trillion-parameter open model; and an OpenRouter dashboard where free stealth models, $0.14/M-token DeepSeek calls, and Claude Sonnet are all one config-line away. Underneath is an unsettled debate about what counts as an 'agent' in the first place — and whether the next decade's infrastructure default is something simpler than what 2024 told us to buy.
Anthropic's Barry Zhang gave the cleanest framing of the year at the AI Engineer Summit, distilled by Shelly Palmer in How Anthropic Thinks About Agents, Workflows, and Tasks (April 2026): a Task is one model call, a Workflow is multiple calls in a control flow you wrote, and an Agent is a model using tools in a loop and choosing its own trajectory. The whole tradeoff space collapses to one structural question: who owns the plumbing, you or the model?
Zhang's economic guardrail is sharper than most public guidance: a 10¢-per-task budget buys roughly 30,000–50,000 tokens, which is workflow territory, not agent territory. A support operation handling 1M tickets a month at 5x necessary token spend burns ~$1.5M/year in waste. His pre-deployment checklist for any team shipping agents: audit every internal project labeled 'agent' against the task/workflow/agent definition, instrument p50 and p95 token costs in production, and stand up trajectory replay before authorizing deployment.
Microsoft's AI Agent Orchestration Patterns guide arrives at the same conclusion from the other direction with a three-rung complexity ladder — direct call → single agent with tools → multi-agent — and a strong prior to climb only when prompt complexity, tool overload, or security boundaries genuinely demand it. The under-appreciated note: sequential orchestration is a bad fit when stages are embarrassingly parallel or when the workflow needs backtracking, two failure modes that actually describe most real coding tasks. Additional info: the same 'producing an answer and evaluating it are different jobs' insight shows up in a Hugging Face forum thread (March 2026) on Thinker/Observer architectures, which the response thread maps cleanly onto existing generator-verifier and process-supervision literature like the Training Verifiers for Math Word Problems paper.
The verification-loop pattern isn't theoretical. SpecLoop (March 2026) is a working example from hardware: an agent generates a spec from RTL, reconstructs RTL from the spec, runs formal equivalence checking, and feeds counterexamples back as a refinement signal. It significantly outperforms LLM-only baselines across multiple models and benchmarks — exactly the 'critic in the loop' that Zhang's taxonomy predicts you'll need anywhere errors compound across iterations.
The clearest concrete instantiation of Zhang's taxonomy is the oh-my-opencode-slim plugin (3.6k stars), which turns OpenCode into a hub-and-spoke 'Pantheon': an Orchestrator delegates to specialists — Explorer (read-only grep), Librarian (docs/web research), Oracle (high-judgment reasoning), Designer (UI), Fixer (scoped patches), and a Background Manager that runs fire-and-forget deep-research jobs. The default preset assigns gpt-5.5 to the high-judgment roles and gpt-5.4-mini to the cheap, scoped ones, and ships with tmux integration that auto-spawns a pane per agent so you can literally watch the team work.
The role-specific prompts are tighter than most production systems'. Looking at the Sisyphus orchestrator source, the persona is blunt — 'SF Bay Area engineer. Work, delegate, verify, ship. No AI slop.' — and the architecture leads with a Phase 0 'Intent Gate' that verbalizes intent before classification, with an explicit Surface-Form-to-True-Intent map ('look into X' → investigation; 'what do you think' → wait for confirmation, don't implement). Per-model prompt builders override delegation behavior for Claude Opus 4.7, GPT-5.4/5.5, Gemini, and Kimi K2.6, with a hard rule never to start implementation unless the user explicitly requests it. The Explorer prompt — quoted in Fabio Biffi's walkthrough (April 2026) — locks the agent to a strict <results><files><answer> XML output format and explicit READ-ONLY constraint.
Biffi's contribution is a one-file JSON config that points every hero at opencode/big-pickle, a free 'stealth' model OpenCode runs during a feedback-gathering window. The result is a fully functional multi-agent dev environment at zero cost — with the explicit warning, in his own words: 'when the product is free, the product might be you', so don't point it at company code. The project itself was renamed to 'oh my OpenAgent' mid-stream, and Z.ai's ZCode changelog shows the same pattern arriving in commercial tooling — six releases between April 22–28, 2026, culminating in v1.6.0's full GUI for creating and managing sub-agents.
The verification ritual is worth stealing even if you don't adopt the plugin: after install, type 'ping all agents' and confirm each specialist responds with its declared model. It's a one-line smoke test for the entire mixed-provider configuration, and it surfaces the kind of silent-fallback misconfiguration that quietly turns your $30/month preset into a $300/month one.
The pricing intelligence behind the per-agent model assignment is unusually skewed in 2026. DeepSeek V4 is the headline: deepseek-v4-flash at $0.14/M input (cache-miss) and $0.28/M output, with cache-hit input pricing reduced to 1/10th of launch (effective April 26, 2026). The Pro tier sits at a 75%-off promotional $0.435/$0.87 per M tokens through May 31, 2026, with a 1M context window and 384K max output. Notably, DeepSeek exposes an Anthropic-format base URL at api.deepseek.com/anthropic, making it a near drop-in replacement in any Claude API client — including the per-agent slots in oh-my-opencode-slim.
For shop-around behavior, OpenRouter remains the default gateway: a free tier with 25+ free models capped at 50 reqs/day, a PAYG tier charging a 5.5% platform fee on top of provider pricing across 300+ models and 60+ providers, and BYOK that's free for the first 1M requests/month before stepping up to 5%. Prompt caching is a first-class feature, which materially changes the economics of the long, repetitive system prompts that agent loops generate. The AnyAPI buyer's guide (March 2026, with the usual self-promotional bias noted) ranks Gemini 2.5 Flash Lite ($0.07/$0.30, 2M context) as the RAG/document winner, Llama 3.3 8B on Groq ($0.05/$0.08) as the speed king for voice/chat, and DeepSeek V3 as the price-per-quality leader for coding.
The expensive end of the market tells a different story. Anthropic's profile reports 300k+ enterprise customers and $5B+ ARR by August 2025, a roughly 57x jump from ~$87M ARR at the start of 2024, with 80% of revenue coming from enterprise. Sonnet 4.5 is positioned as the coding/agents tier, Haiku 4.5 for cost-sensitive work, and Opus 4.1 for complex reasoning. The customer outcomes Anthropic markets are unusually concrete: Doctolib onboarding compressed from weeks to days, Classmethod's code review cycle from 24 hours to 1 hour, and Novo Nordisk's clinical study reports from 10+ weeks to 10 minutes.
The practical synthesis for a multi-agent setup: route Orchestrator and Oracle to Claude Sonnet 4.5 or DeepSeek V4 Pro; pin Explorer/Fixer/Librarian to DeepSeek V4 Flash or Gemini Flash Lite; keep a Haiku fallback for anything customer-facing where nuance and refusal-behavior matter; and watch your p95 token cost like a hawk because mixed-provider setups make cost regressions invisible until the bill arrives.
The most informative real-world report on local frontier inference is spicyneuron's six-month writeup (April 2026) of running 600B+ parameter models on a 512GB Mac Studio M3. The headline numbers: simple chat replies in seconds, ~30 seconds for a 7,000-token edit, and ~90 seconds to run Claude Code with a ~16,000-token system prompt — slower than APIs, but streaming at 3x reading speed. The author's bluntest finding: open-weight models in his hands lag frontier APIs by 6 to 12 months, regardless of what benchmarks claim.
Two practical tweaks matter. First, macOS caps GPU memory at 75% of system RAM by default, which on a 512GB box wastes ~120GB; raising iogpu.wired_limit_mb via a Launch Daemon (sysctl alone doesn't persist across reboots, and modern macOS silently ignores /etc/sysctl.conf) reclaims it. Second, in 2026 MLX is consistently 10–25% faster than llama.cpp on Apple Silicon, and MoE architectures are the architectural sweet spot — full weights stay resident in unified memory while only active params hit inference cost, which is exactly how Kimi K2.5's 1T parameters become tractable on a desktop.
If one Mac isn't enough, exo (44.2k stars) clusters multiple devices with day-0 RDMA over Thunderbolt 5 (~99% latency reduction) and tensor parallelism that scales to 1.8x on 2 devices and 3.2x on 4, demoed on 4× M3 Ultra Mac Studios running DeepSeek V3.1 (8-bit) and Kimi-K2-Thinking (4-bit). It exposes OpenAI Chat Completions, Claude Messages, OpenAI Responses, and Ollama-compatible APIs simultaneously, with recent commits adding DeepSeek V4 Flash/Pro and Kimi K2.6. The architectural reasoning is laid out in Apple Silicon vs NVIDIA CUDA (August 2025): unified memory at 546 GB/s and 40–80W vs CUDA's 1 TB/s VRAM at 300–700W — train on CUDA, infer/prototype on Apple Silicon.
The CUDA-side counter-argument is the Additional info: Supermicro 4U GPU SuperServer AS-4125GS-TNRT2 (date unavailable): up to 10 double-width PCIe 5.0 GPUs (H100 NVL, RTX PRO 6000 Blackwell, A100, L40S, MI210), dual EPYC 9005/9004 (up to 192 cores / 384 threads), 6TB DDR5, optional NVLink Bridge or AMD Infinity Fabric Link. It's the right answer if you're training, the wrong answer if you're trying to run a coding-agent pantheon at your desk without rewiring the room. And on the model-selection side, Qwen3.6-27B (April 2026) is being marketed as a 27B model that beats a 397B model on coding — the kind of size-to-capability ratio that makes it a plausible default for the cheap roles (Explorer, Fixer) on a single Mac Studio.
David Linthicum's Enterprises are rethinking Kubernetes (April 2026) makes the harder argument: K8s isn't going away, but it's becoming plumbing rather than a strategic buying decision. The operational tax — multiplying clusters, sprawling toolchains, risky upgrades, governance as its own discipline — turned out to be the actual cost, while the headline benefit (portability) was undermined by ecosystem dependencies in storage, identity, and observability that produced soft lock-in regardless. Platform-engineering teams increasingly hide K8s behind internal developer platforms (IDPs), which is exactly the 'don't build agents, build workflows' move applied to infrastructure.
Linthicum's caveat is important: K8s is still the right answer for digital-natives, regulated multicloud workloads, and highly customized environments. The point isn't that it was wrong — it's that the default-everywhere assumption was wrong, and enterprise architects are catching up to that. The pattern rhymes with the agent-vs-workflow debate: both stories are about resisting the urge to deploy maximum complexity when a simpler abstraction would do the job at a fraction of the operational cost.
The financial-skeptic counterweight comes from Additional info: Ed Zitron's How OpenAI ends and takes Oracle with it (date unavailable), a 40-minute bear case tying Oracle's fortunes specifically to OpenAI's compute commitments. Whether or not you buy the thesis, it's a useful pairing with the Anthropic enterprise numbers — $5B+ ARR, 300k+ enterprise customers — because it forces the question of how much of the current AI infrastructure spend is durable demand versus circular financing between hyperscalers and a handful of model labs.
For the supporting infrastructure layer: tmux (latest 3.6a) is the unsexy primitive that makes the 'watch your agents work in real-time' UX possible — oh-my-opencode-slim's auto-spawned panes are just tmux underneath. And Tailscale, with its Aperture beta positioned for AI workloads, is the natural connectivity glue for distributed local-AI setups (exo nodes across locations, BYOK API access from a laptop into a Mac Studio at home). Both are reminders that a lot of the 2026 'AI stack' is the same boring, durable infrastructure dressed up in new use cases.
The through-line across all five clusters is the same managerial instinct: match complexity to value, then verify before scaling. Workflow before agent, single Mac before cluster, DeepSeek Flash before Sonnet, IDP before raw Kubernetes. The teams that ship in the next year will be the ones that internalize Zhang's checklist — instrument p50/p95 token cost, audit your 'agents' for whether they're actually workflows in costume, and stand up trajectory replay before deployment — and treat every additional layer of orchestration as a debt to be justified, not a default to be assumed.