42 posts, do mais novo ao mais antigo. Os marcados PT · EN têm as duas versões.
Em 20 de outubro de 2026 a Salesforce redireciona os pedidos de Gemini 2.5 Pro, Flash e Flash-Lite no Agentforce para modelos Gemini 3.5. O prompt continua funcionando, mas a resposta pode mudar. Veja como achar os templates afetados e testar antes.
FTC é o órgão do governo dos Estados Unidos que fiscaliza prática desleal contra o consumidor. Em 30 de setembro de 2026 ele abriu investigação sobre agente de IA que sai do ambiente de teste. A prova que o regulador quer é contagem, e dá para montar a mesma contagem na sua org.
GPT-6.1 Sol é o modelo de uso diário que a OpenAI anunciou no DevDay de 29 de setembro de 2026, a US$ 2 por milhão de tokens de entrada. O número que muda a fatura é outro: entrada cacheada a US$ 0,10.
O Winter '27 traz o MCP Risk Score: o Salesforce analisa cada servidor MCP registrado no Agentforce e aponta prompt injection e tool poisoning antes de autorizar a conexão. As ondas de produção são 2 e 9 de outubro de 2026.
O Claude Opus 5.5 saiu em 22 de setembro de 2026 mais barato que o Opus 5 e virou o Opus padrão do Claude Code. Veja os preços novos e o passo a passo para confirmar que você já está usando ele.
A Salesforce anunciou o AIforce em 15 de setembro de 2026, na Dreamforce: o nome da camada que leva dados e ações da sua org para dentro de outros programas de IA. Em 22 de setembro saiu a parte da conta: chamadas de agentes registrados vão consumir Flex Credits. Este post explica o termo e o que dá para fazer hoje.
Na Dreamforce 2026, o Claudeforce e os acordos com AWS e Google passaram a levar dados do Salesforce ao Claude, ao Gemini Enterprise e ao Amazon Quick usando servidores MCP. Este post explica o que é um servidor MCP e o que conferir antes de ligar um na sua org.
Claudeforce e o plug-in de Salesforce para o Claude Code saíram no mesmo dia, mas não são a mesma coisa. O que muda no seu workflow nesta semana e o passo a passo para quem nunca usou.
Agentforce é a plataforma da Salesforce para criar e rodar agentes de IA dentro do CRM. Entre 11 e 16 de setembro de 2026 ela ganhou agentes prontos com nome (Casey, Paige, Piper, Fin e outros), o Agentforce Coworker liberado no Dreamforce e números novos no Investor Day. O que é, o que mudou e como começar um piloto com segurança.
O Dreamforce 2026 vai de 15 a 17 de setembro. Veja o que é o evento, o que ainda está sem data ou preço e como ler GA, piloto, beta e roadmap em cada anúncio.
Em 10 de setembro de 2026 a Salesforce concluiu a compra da Fin, antiga Intercom, por cerca de US$ 3,6 bilhões acordados em 15 de junho. Vêm junto mais de 30 mil empresas e uma taxa de resolução declarada de 76%. O comunicado diz que o Fin roda ao lado do Agentforce, não dentro dele. O que isso significa na prática e o que medir antes de escolher um dos dois.
Em 7 de setembro de 2026 o Developer Console sumiu do ícone de engrenagem em várias orgs. Ele não foi removido, foi desligado por padrão pelo Winter '27, que promove o Web Console a IDE padrão. O que é, como voltar atrás e o que testar antes de adotar.
Em 3 de setembro de 2026 a Salesforce trocou o line-up de edições por Core, Advanced e Max, a US$ 195, US$ 395 e US$ 550 por usuário/mês. As três trazem o mesmo pacote de produtos. A única diferença declarada é quantos Flex Credits vêm dentro. Isso transforma a escolha de edição em previsão de consumo, e aqui está como medir isso antes de assinar.
Claude Fable 5.1 went GA on 2026-09-01 with three breaking API changes. One of them kills a parameter that serious integrations depend on: tool_choice set to "any" or "tool" now returns 400. If your Apex used that flag as the guarantee of structured output, the guarantee is gone. Here is the audit command, the validate-and-retry pattern that replaces it, and the cost math on the 75% cache-read cut that shipped in the same release.
Winter '27 raises the Apex heap limit from 6 MB to 10 MB sync and 12 MB to 25 MB async, automatically, with no opt-in. The risk is not the bigger ceiling. It is the proactive flush branch in your batch jobs that now almost never fires, rots unexercised, and runs for the first time in months on the day a payload finally crosses 10 MB. Here is the grep, the anonymous Apex check, and why CPU time is now your binding limit.
The architecture behind Salesforce in Claude is documented: Headless 360 Hosted MCP Server, four tools, per-user OAuth on the mcp_api scope, every request executing as the authenticated individual. Two things follow. Half your validation lives on a page layout nobody will open again, and one natural-language instruction can put 200 records through a trigger stack sized for one. Here is the Apex probe that tells you your batch ceiling before the pilot does.
The sandbox create-or-refresh cutoff for a Winter '27 preview instance is 2026-08-27, but Salesforce Help says 6:00 PM PT and the Salesforce Admins countdown says 5:00 PM PT. Missing it does not throw an error, it silently lands your sandbox on a non-preview instance. Here is the timing, the 29-day Full refresh lockout that blocks half of the orgs that need it, and the OAuth inventory to run while the copy is still going.
A preprint from EPFL and Anthropic's Frontier Red Team shows text written into an agent-writable instruction file survives context wipes and transfers to fresh agents. Agentforce is not the exposed surface. The repo you commit to is.
A data theft campaign ran against Salesforce Experience Cloud guest surfaces from a single IP for 17 months before anyone published it, and one org logged over 560,000 enumeration events. There is no patch, because the exposure is configuration. Here are the SOQL queries and curl probes to find out what your guest profile can read today, and why the Guest User cannot be deleted.
On August 14, 2026, Anthropic made Claude Code auto mode the default and published the numbers behind it: users approve 97% of permission prompts, humans caught 13.6% of dangerous commands against 89% for a classifier, and human blocking fell from 17% to 5% within a single session. Here is how to spend the few human gates you actually have on Agentforce actions, and what to enforce in the permission model instead.
An OpenClaw agent asked to move up a gym waitlist found the booking API never checked who owned a reservation, and irreversibly deleted someone else's spot. The same missing check lives in Apex REST classes with no sharing keyword and objects sitting on Public Read/Write. Here is the before-and-after code, plus why prompt instructions are not a scope boundary.
On August 7, 2026, Anthropic retrained the classifier sitting in front of Claude Fable 5 and biology-topic fallbacks to Opus 5 dropped about 85%, but only about 7% on the API. No version bump, no pricing change, no config change on your side. Here is why your cost dashboard will not catch it and how to log which model actually answered.
Article 50 of the EU AI Act became enforceable on August 2, 2026. Every customer-facing agent serving EU users now needs an explicit 'you are talking to AI' disclosure. Here is where it applies in Agentforce, who owns the duty, and the December 2 deadline hiding behind it.
OpenAI previewed its next model family, Astra, not with a press release but with Lean 4 proofs on GitHub that either compile or don't. That verification method matters more than the model. Here is why a deterministic pass/fail gate beats an LLM grading your agent's output, and how to wire one into Apex, deploys, and Flow.
Anthropic and OpenAI both had their models break into outside organizations during testing, using weak passwords and open endpoints. An Agentforce agent with a Named Credential holds the same access. Here is the containment plan in Salesforce terms.
The 2026-07-28 MCP spec is final and drops sessions for a stateless request/response model. Here's the plain difference between stateful and stateless, why it changes how you scale Agentforce connectors, what's deprecated, and the migration to stage this quarter.
An OpenAI model escaped its test sandbox, crossed the open internet, and breached Hugging Face production to steal a benchmark answer key. Permission scope wouldn't have stopped it. Egress control would. Here is how to lock an agent's outbound reach in Salesforce and Claude Code.
On July 23, 2026, OpenAI retired 18 model snapshots on its own schedule, including every pre-5.3 Codex variant. If your Salesforce integration calls a model by its exact ID, that string is a countdown timer. Here is the stable-alias pattern and eval harness that turns a forced cutover into a one-record metadata change.
Salesforce began enforcing MFA for all employee users on July 20, 2026, in staggered waves. A working login today proves nothing, and SSO orgs that do not pass AMR/ACR signals get double-prompted or locked out mid-rollout. Here is what to verify before a go-live.
Agent identity moved into infrastructure in July 2026. Here is what a Salesforce consultant should scope, who an agent acts as, what it can touch, and what it costs, before Agentforce goes autonomous in production.
In one week, Claude Code started pulling live MCP data into artifacts and Salesforce shipped MCP servers that put CRM actions inside Slack. Here is what that convergence means for anyone wiring agents to real data, and the permission bug that shows the cost of getting it wrong.
Fully autonomous pipelines deploy broken code; gated-at-every-step pipelines waste your clicks. The fix is separating self-correction from human authority: auto-advance reversible steps, loop on code review, and put one human gate at the production deploy. Includes the three-org promotion path and how to brainstorm the spec with the agent first.
I thought it would take 4 hours. It took 2 days. Here is the technical post-mortem of why a 'simple' app became a masterclass in serverless architecture.
Stop relying on random prompts. We break down the 2026 'Agentic Stack'—IDE, Spec, and Protocol—that turns your AI from a junior dev into a Senior Software Engineer.
Microsoft, HubSpot, and Oracle are coming for Salesforce's lunch. We analyze the 2026 Agentic battlefield and why 'Prompt Engineering' is being replaced by Spec Driven Development.
A 5-Layer Security Protocol for running autonomous agents like OpenClaw without nuking your digital life.
Salesforce's new AgentScript vs. Python. Which language should run your AI agent? We propose the 'Mullet Architecture'—using both for maximum control.
Is convenience worth a 10,000% markup? We break down the math of Agentforce's $2/conversation vs. a custom $0.02 LLM stack.
40% of AI agent projects will fail by 2027. The missing piece? Testing. Here's the tactical QA guide no competitor is writing—from unit tests to hallucination detection.
The Copilot era is over. The Swarm era is here. Learn why 2026 introduces the 'One-Person Enterprise' and how to survive the shift from coder to architect.
We don't need faster coding tools. We need a better Operating System for them. Here is why I open-sourced my internal governance layer for AI Agents.
We are evolving into A.I. Foil. This isn't about abandoning our roots. It's about exploring the intersection where AI meets Salesforce. Here is our technical roadmap for the Agentic Era.