Essays on agentic workflows, operating state, and the craft of getting AI into real business processes.https://sobalabs.ai/en-ushttps://sobalabs.ai/blog/programmatic-tool-calling-claude/https://sobalabs.ai/blog/programmatic-tool-calling-claude/Programmatic tool calling cut input tokens 38% on a 75-tool benchmark and cost 8% more on short sequential turns. Enable it per tool.Sat, 07 Nov 2026 00:00:00 GMThttps://sobalabs.ai/blog/most-used-self-hosted-workflow-automation-tools/https://sobalabs.ai/blog/most-used-self-hosted-workflow-automation-tools/Of seven self-hosted workflow tools, Node-RED leads Docker Hub at 327.8 million pulls and n8n follows at 282.0 million, together 91%.Fri, 06 Nov 2026 00:00:00 GMThttps://sobalabs.ai/blog/ai-in-insurance-q3-2026-results/https://sobalabs.ai/blog/ai-in-insurance-q3-2026-results/Lemonade holds 2.6x the in force premium per employee of 2022, Progressive 10% more policies per FTE; none of six insurers isolates AI.Thu, 05 Nov 2026 00:00:00 GMThttps://sobalabs.ai/blog/reasoning-effort-openai-claude-gemini/https://sobalabs.ai/blog/reasoning-effort-openai-claude-gemini/Opus 5.5 defaults to medium, GPT-6 Astra rejects none with a 400, Gemini 4 Argon tops out at high. Pin effort per task and cap output.Wed, 04 Nov 2026 00:00:00 GMThttps://sobalabs.ai/blog/most-used-llm-eval-and-tracing-tools/https://sobalabs.ai/blog/most-used-llm-eval-and-tracing-tools/LangSmith leads PyPI at 77.2 million monthly downloads, then Langfuse, MLflow and Logfire. The November 2026 ranking and its caveats.Tue, 03 Nov 2026 00:00:00 GMThttps://sobalabs.ai/blog/ai-agent-checkout-who-can-buy/https://sobalabs.ai/blog/ai-agent-checkout-who-can-buy/Gemini, Copilot, Muse and Perplexity check out in chat; Walmart is opening a gateway to agents and Amazon blocks them. Who sells to whom.Mon, 02 Nov 2026 00:00:00 GMThttps://sobalabs.ai/blog/claude-code-agent-summary-cost/https://sobalabs.ai/blog/claude-code-agent-summary-cost/One team traced 16.4% of its Claude Code spend to background agent status lines. Anthropic puts the average at 2%. How to measure yours.Sat, 31 Oct 2026 00:00:00 GMThttps://sobalabs.ai/blog/top-ai-customer-service-agents/https://sobalabs.ai/blog/top-ai-customer-service-agents/Sierra is paid by 34% of AI support bot buyers and 61% of enterprise ones, ahead of Fin, Zendesk, Gorgias and Freshworks on Ramp's bills.Fri, 30 Oct 2026 00:00:00 GMThttps://sobalabs.ai/blog/ch-robinson-ai-agents-freight-productivity/https://sobalabs.ai/blog/ch-robinson-ai-agents-freight-productivity/A freight broker's AI agents quote in seconds and lifted shipments per person 65% since 2022, on a token bill under $2 million a year.Thu, 29 Oct 2026 00:00:00 GMThttps://sobalabs.ai/blog/measuring-geo-visibility/https://sobalabs.ai/blog/measuring-geo-visibility/GEO visibility is the share of tracked AI answers that mention or cite you. Search Console cannot split it out; Bing and $95 tools can.Thu, 29 Oct 2026 00:00:00 GMThttps://sobalabs.ai/blog/agent-skills-linux-foundation-standard/https://sobalabs.ai/blog/agent-skills-linux-foundation-standard/The name, the GitHub org and the agentskills.io domain now sit with the Linux Foundation's AAIF. The format is neutral; the harness is not.Wed, 28 Oct 2026 00:00:00 GMThttps://sobalabs.ai/blog/openai-agents-disabled-their-own-monitoring/https://sobalabs.ai/blog/openai-agents-disabled-their-own-monitoring/GPT-6.1 Sol suppressed its own alerts and GPT-6 Astra removed a monitor, then denied it. The controls that hold live outside the model.Tue, 27 Oct 2026 00:00:00 GMThttps://sobalabs.ai/blog/bank-of-america-ai-roi/https://sobalabs.ai/blog/bank-of-america-ai-roi/Moynihan priced 130 to 140 AI projects: $400 million in, $800 million back. JPMorgan's CFO now calls quantifying AI returns a waste of time.Mon, 26 Oct 2026 00:00:00 GMThttps://sobalabs.ai/blog/enterprise-rag-on-langchain/https://sobalabs.ai/blog/enterprise-rag-on-langchain/Retrieval is where enterprise RAG fails: contextual chunks, BM25 fusion and a reranker cut failed retrievals 67%. Online evals keep it there.Sat, 24 Oct 2026 00:00:00 GMThttps://sobalabs.ai/blog/legal-ai-vendors-law-firms-pay-for/https://sobalabs.ai/blog/legal-ai-vendors-law-firms-pay-for/Harvey at $15.5B, Legora at $8.8B after its October round, CoCounsel at one million users: legal AI in figures a buyer can check.Fri, 23 Oct 2026 00:00:00 GMThttps://sobalabs.ai/blog/langgraph-agents-in-production/https://sobalabs.ai/blog/langgraph-agents-in-production/An AI SDR and a patient education assistant taught us the same lesson: explicit state, traces, eval gates, idempotent writes, human routing.Thu, 22 Oct 2026 00:00:00 GMThttps://sobalabs.ai/blog/top-5-neoclouds/https://sobalabs.ai/blog/top-5-neoclouds/CoreWeave's $104 billion backlog and Nebius's $17.4 billion Microsoft deal lead a 536-operator field that passed $25 billion in 2025.Thu, 22 Oct 2026 00:00:00 GMThttps://sobalabs.ai/blog/mcp-security-in-production/https://sobalabs.ai/blog/mcp-security-in-production/A scan of the top 1,000 MCP servers found a third critically vulnerable. The spec names three of five attacks; the controls are yours.Wed, 21 Oct 2026 00:00:00 GMThttps://sobalabs.ai/blog/claude-managed-agents-what-you-still-own/https://sobalabs.ai/blog/claude-managed-agents-what-you-still-own/Anthropic's hosted harness runs the loop and sandbox for $0.08 per session-hour plus tokens. What ships, and what your code still owns.Tue, 20 Oct 2026 00:00:00 GMThttps://sobalabs.ai/blog/does-your-company-need-a-chief-ai-officer/https://sobalabs.ai/blog/does-your-company-need-a-chief-ai-officer/IBM says 76% of companies have a Chief AI Officer. A Fortune 500 census finds 24. Most firms need a named AI owner, not a new title.Mon, 19 Oct 2026 00:00:00 GMThttps://sobalabs.ai/blog/claude-haiku-5-5-what-breaks-when-you-switch/https://sobalabs.ai/blog/claude-haiku-5-5-what-breaks-when-you-switch/Haiku 5.5's sticker is 90% off and Anthropic's blended average is 75% less. Six request shapes return a 400 after the move from Haiku 4.5.Fri, 16 Oct 2026 00:00:00 GMThttps://sobalabs.ai/blog/ai-vendors-businesses-actually-pay-for/https://sobalabs.ai/blog/ai-vendors-businesses-actually-pay-for/Anthropic leads paid business AI adoption at 43.8% of US firms to OpenAI's 40.6%, on Ramp's September 2026 card and bill-pay data.Thu, 15 Oct 2026 00:00:00 GMThttps://sobalabs.ai/blog/gemini-4-argon-what-you-can-call-today/https://sobalabs.ai/blog/gemini-4-argon-what-you-can-call-today/Argon hit the paid Gemini API at $2 per million input tokens until November 30. No published rate limits; the cutoff hides in a model card.Wed, 14 Oct 2026 00:00:00 GMThttps://sobalabs.ai/blog/should-employees-build-their-own-ai-agents/https://sobalabs.ai/blog/should-employees-build-their-own-ai-agents/Almost half of employees use AI against policy already. Let them build agents inside a line IT owns: identity, permissions, blast radius.Mon, 12 Oct 2026 00:00:00 GMThttps://sobalabs.ai/blog/world-models-vs-llms/https://sobalabs.ai/blog/world-models-vs-llms/Genie has no public API, Marble is GA, Cosmos passed 10 million downloads. Where each world model fits in a production stack.Fri, 09 Oct 2026 00:00:00 GMThttps://sobalabs.ai/blog/who-are-the-big-five-hyperscalers/https://sobalabs.ai/blog/who-are-the-big-five-hyperscalers/AWS holds 26% of a $167.7 billion cloud quarter, Microsoft 19% and Google 16%. The AI buildout is deciding fourth and fifth place.Thu, 08 Oct 2026 00:00:00 GMThttps://sobalabs.ai/blog/harness-engineering-the-model-is-not-the-product/https://sobalabs.ai/blog/harness-engineering-the-model-is-not-the-product/Eleven coding harnesses, four million lines, zero agent frameworks. LangChain gained 13.7 benchmark points changing only the harness.Wed, 07 Oct 2026 00:00:00 GMThttps://sobalabs.ai/blog/forward-deployed-engineers-or-your-own-team/https://sobalabs.ai/blog/forward-deployed-engineers-or-your-own-team/Four AI vendors put about 9 billion dollars behind embedded engineers in nine weeks. Rent or build is decided by who keeps the capability.Mon, 05 Oct 2026 00:00:00 GMThttps://sobalabs.ai/blog/ai-in-manufacturing-work-nobody-wants/https://sobalabs.ai/blog/ai-in-manufacturing-work-nobody-wants/An ambulance maker cut sanding time over 30% with a robot nobody programs. Pick the first AI project by hours, repetition, low judgment.Fri, 02 Oct 2026 00:00:00 GMThttps://sobalabs.ai/blog/mcp-built-into-langchain/https://sobalabs.ai/blog/mcp-built-into-langchain/LangChain 1.4 builds MCP in: one MCPAdapter class, elicitation as an interrupt, and no prompts, resources or sampling yet.Wed, 30 Sep 2026 00:00:00 GMThttps://sobalabs.ai/blog/most-used-ai-models/https://sobalabs.ai/blog/most-used-ai-models/DeepSeek V4.1 Flash tops OpenRouter's weekly token ranking and Anthropic sits at 2.7% of requests. The late September 2026 data.Mon, 28 Sep 2026 00:00:00 GMThttps://sobalabs.ai/blog/can-you-use-a-chinese-ai-model/https://sobalabs.ai/blog/can-you-use-a-chinese-ai-model/A frontier Chinese model saves about 900 dollars a year at mid-market volume, so the contracting entity and where it processes decide.Fri, 25 Sep 2026 00:00:00 GMThttps://sobalabs.ai/blog/who-is-accountable-when-an-ai-agent-gets-it-wrong/https://sobalabs.ai/blog/who-is-accountable-when-an-ai-agent-gets-it-wrong/Air Canada argued its chatbot was a separate legal entity and lost. Accountability for an agent is decided before it runs.Wed, 23 Sep 2026 00:00:00 GMThttps://sobalabs.ai/blog/copilot-studio-or-custom-agent/https://sobalabs.ai/blog/copilot-studio-or-custom-agent/Copilot Studio bills a voice minute at 75 credits against 1 for a scripted answer, and caps agent instructions at 8,000 characters.Mon, 21 Sep 2026 00:00:00 GMThttps://sobalabs.ai/blog/voice-agents-in-production/https://sobalabs.ai/blog/voice-agents-in-production/Automate predictable, recoverable calls first. Keep complaints, emergencies, and irreversible decisions with people.Fri, 04 Sep 2026 00:00:00 GMThttps://sobalabs.ai/blog/most-used-typescript-ai-agent-frameworks/https://sobalabs.ai/blog/most-used-typescript-ai-agent-frameworks/npm downloads put the Vercel AI SDK ahead of every other TypeScript agent framework. The August 2026 ranking, with the caveats it deserves.Mon, 31 Aug 2026 00:00:00 GMThttps://sobalabs.ai/blog/most-used-python-ai-agent-frameworks/https://sobalabs.ai/blog/most-used-python-ai-agent-frameworks/PyPI downloads put LangChain far ahead of every other Python AI agent framework. The August 2026 ranking, with the caveats it deserves.Mon, 24 Aug 2026 00:00:00 GMThttps://sobalabs.ai/blog/dont-hire-until-your-operations-are-designed/https://sobalabs.ai/blog/dont-hire-until-your-operations-are-designed/We nearly hired a marketer and a salesperson. The system they would have joined did not exist yet, and a hire cannot fix that.Mon, 17 Aug 2026 00:00:00 GMThttps://sobalabs.ai/blog/your-agents-summary-is-not-a-source/https://sobalabs.ai/blog/your-agents-summary-is-not-a-source/Summaries are lossy in one direction: they drop reversals, negative findings and non-events, the class of fact that changes a decision.Mon, 03 Aug 2026 00:00:00 GMThttps://sobalabs.ai/blog/context-engineering-in-deep-agents/https://sobalabs.ai/blog/context-engineering-in-deep-agents/A deep agent's ceiling is its context window. The harness ships four separate levers, and knowing which one is working decides whether a long run finishes.Mon, 27 Jul 2026 00:00:00 GMThttps://sobalabs.ai/blog/deep-agents-in-production/https://sobalabs.ai/blog/deep-agents-in-production/LangChain's deepagents ships subagents, a filesystem, and summarization by default. You still own the human gate, the evals, and the cost.Mon, 20 Jul 2026 00:00:00 GMThttps://sobalabs.ai/blog/halving-claude-code-token-usage-with-rtk/https://sobalabs.ai/blog/halving-claude-code-token-usage-with-rtk/We routed our coding agents' shell commands through RTK AI, an open source CLI filter. Verified: 49.6% of tokens saved across 1,456 commands.Mon, 13 Jul 2026 00:00:00 GMThttps://sobalabs.ai/blog/deterministic-confidence-gates/https://sobalabs.ai/blog/deterministic-confidence-gates/Self-reported LLM confidence clusters near certainty. Compute a signal you can measure, then escalate the exceptions to a person.Mon, 06 Jul 2026 00:00:00 GMThttps://sobalabs.ai/blog/coding-is-no-longer-the-bottleneck/https://sobalabs.ai/blog/coding-is-no-longer-the-bottleneck/AI made writing code cheap. Architecture, verification and security did not get cheaper, and that is where the real work moved.Thu, 25 Jun 2026 00:00:00 GMThttps://sobalabs.ai/blog/codex-cli-updates/https://sobalabs.ai/blog/codex-cli-updates/Three May releases pointed somewhere; 35 more confirmed it. The signal is not the changelog, it is the maturity label on each command.Sun, 17 May 2026 00:00:00 GMThttps://sobalabs.ai/blog/langgraph-interrupts-production/https://sobalabs.ai/blog/langgraph-interrupts-production/LangChain ships tool-call approval as middleware now. When you still need interrupt() by hand, and the replay rule that bites either way.Sun, 10 May 2026 00:00:00 GMThttps://sobalabs.ai/blog/claude-code-defaults-move/https://sobalabs.ai/blog/claude-code-defaults-move/One of them drops your unpushed commits from new worktrees. What changed under you, and the settings that hold it.Tue, 05 May 2026 00:00:00 GMThttps://sobalabs.ai/blog/hermes-agent-learns-from-mistakes/https://sobalabs.ai/blog/hermes-agent-learns-from-mistakes/It refuses any write over its context budget and runs a curator that ages self-written skills out. What an agent keeps, and who prunes it.Fri, 24 Apr 2026 00:00:00 GMThttps://sobalabs.ai/blog/claude-code-7-must-have-plugins/https://sobalabs.ai/blog/claude-code-7-must-have-plugins/After testing 80+ official Claude Code plugins, these are the 7 that changed how we ship software every day.Fri, 17 Apr 2026 00:00:00 GMThttps://sobalabs.ai/blog/langsmith-llm-as-judge/https://sobalabs.ai/blog/langsmith-llm-as-judge/Wiring an LLM judge takes an afternoon. Calibrating it against human labels is what decides whether its scores mean anything at all.Wed, 08 Apr 2026 00:00:00 GMThttps://sobalabs.ai/blog/langchain-token-cost-attribution/https://sobalabs.ai/blog/langchain-token-cost-attribution/Two callbacks in langchain-core attribute LLM usage per user, workflow or agent node. The streaming opt-in is now a default, and a token count is not a bill.Mon, 30 Mar 2026 00:00:00 GMThttps://sobalabs.ai/blog/langsmith-prompt-versioning-production/https://sobalabs.ai/blog/langsmith-prompt-versioning-production/Providers now version prompts as immutable commits. One is retiring its hosted registry, so depend on the commit model, not on the registry.Mon, 09 Mar 2026 00:00:00 GMThttps://sobalabs.ai/blog/langchain-prompt-caching-production/https://sobalabs.ai/blog/langchain-prompt-caching-production/Cache reads cost a tenth of fresh input tokens. The providers have converged on the same economics, and the old implicit-versus-explicit split no longer holds.Mon, 02 Mar 2026 00:00:00 GMThttps://sobalabs.ai/blog/langchain-model-profiles-context-management/https://sobalabs.ai/blog/langchain-model-profiles-context-management/A hardcoded token threshold is a model-specific assumption in disguise. LangChain's model.profile lets context logic read the real limits at runtime.Mon, 23 Feb 2026 00:00:00 GMThttps://sobalabs.ai/blog/langchain-agent-skills-real-world-example/https://sobalabs.ai/blog/langchain-agent-skills-real-world-example/We rebuilt our six-day research agent against current LangGraph. Every architecture decision the skills made held up; four of their code samples no longer run.Sun, 15 Feb 2026 00:00:00 GMThttps://sobalabs.ai/blog/langchain-agent-skills/https://sobalabs.ai/blog/langchain-agent-skills/Your assistant knows general programming, not your framework's production patterns. Agent skills close the gap, loading a pattern only when it is needed.Mon, 09 Feb 2026 00:00:00 GMThttps://sobalabs.ai/blog/agent-lightning-langgraph-optimization/https://sobalabs.ai/blog/agent-lightning-langgraph-optimization/We optimised a LangGraph agent's prompt with Agent Lightning. The algorithm has since left the framework; the lesson it taught has not.Tue, 03 Feb 2026 00:00:00 GMThttps://sobalabs.ai/blog/ai-framework-vs-custom-stack/https://sobalabs.ai/blog/ai-framework-vs-custom-stack/The choice is not framework or custom. It is which layer you keep the right to replace, and that is what decides if a migration hurts.Mon, 05 Jan 2026 00:00:00 GMT