Adeilson Brito

Agentic systems treat their prompts as configuration. That assumption is wrong. Agents change behavior when prompts change. But unlike code, prompt changes don't produce error messages. The agent still produces coherent text. It just optimizes for a different target, against a different implicit definition of success. Code changes are subject to CI/CD: unit tests, integration tests, behavioral diffs. Prompt changes have no equivalent. A prompt is not configuration — it is the behavioral definition of the system, and changing it is a production code change with no regression harness. The failure mode is silent: the agent still passes basic coherence checks while drifting toward the wrong objective. Teams catch this only through user reports or incident review, not through any automated signal in the deployment pipeline. The production consequence is that prompt iteration without behavioral testing is a liability, not an optimization. Every prompt change needs a version-controlled record paired with a behavioral diff: a small set of invariant inputs that should produce invariant outputs, and a scoring mechanism that flags divergence. This is the agentic equivalent of unit tests. Without it, production agent behavior degrades silently as prompts are iterated by well-intentioned engineers who have no signal that the system has changed its optimization target. The non-obvious takeaway: the absence of prompt regression testing is not a tooling gap. It is an architectural gap. Agents without behavioral test harnesses are systems deployed without observability — you will know something went wrong, but you will not know when, why, or which prompt change caused it.

@MarioNawfal He's an absolute idiot

🇬🇧 Note to Starmer:

The large language model context window functions like system RAM: without an external orchestration layer, you quickly run into memory fragmentation and state overflow during complex reasoning loops. This architectural reality is exactly why the Model Context Protocol (MCP) is gaining rapid standardization — especially in mature implementations like the Harness MCP server, which cleanly decouples tool execution from the raw prompt payload. When autonomous agents talk directly to unpaginated, highly verbose enterprise APIs, they burn through context limits with redundant data. The result feels like classic system thrashing: the model wastes expensive tokens re-processing bloated JSON instead of advancing its actual reasoning. A standardized dispatch layer changes that. Tool outputs get paginated, filtered, and managed statefully outside the model. Only the precise semantic delta needed for the next decision step flows back into the context window — keeping the reasoning space clean and focused. Engineering teams should move away from hardcoding bespoke, monolithic API integrations directly into their agentic frameworks. Instead, architects should deploy dedicated resource management servers that handle authentication, retries, payload truncation, and state — all outside the generative model’s awareness. Monitoring also needs to evolve. Beyond latency and error rates, site reliability teams should now track context utilization, token eviction frequency, and tool-response bloat as first-class health metrics for autonomous workloads. Ultimately, the operational reliability of an agentic system is limited less by the raw intelligence of the foundational model and more by the infrastructure discipline applied to its context management pipeline. (Source: Harness Engineering Blog / Model Context Protocol Design)

A true Christian life rests on three pillars that defined Jesus himself: radical love, humble service, and deep trust in God. Radical love means loving God with all your heart and loving your neighbor—even your enemies—as yourself. Humble service is following Jesus’ example of washing feet, healing the sick, and putting others first. Deep trust in God is living without worry, seeking His kingdom first, and surrendering your will to His. When the Holy Spirit makes you a new creation, these three stop being goals you chase and become the natural overflow of Christ living in you.

This is genuinely impressive neuroscience + AI work. TRIBE v2 predicting fMRI responses across video, audio, and text from just 500+ hours of data — and open-sourcing the model, code, and demo — is a real win for research. Could seriously speed up brain science and help with neurological conditions. That said… it’s Meta. They already mastered hacking our attention with Reels and feeds. Now they have a model that can predict which pixels and sounds will light up our brains the hardest. The jump from “optimize for engagement” to “optimize for actual neural reward circuits” feels uncomfortably small. Cool tool for science. Mildly terrifying upgrade for the attention economy.

Today we're introducing TRIBE v2 (Trimodal Brain Encoder), a foundation model trained to predict how the human brain responds to almost any sight or sound. Building on our Algonauts 2025 award-winning architecture, TRIBE v2 draws on 500+ hours of fMRI recordings from 700+ people to create a digital twin of neural activity and enable zero-shot predictions for new subjects, languages, and tasks. Try the demo and learn more here: go.meta.me/tribe2

@pierjeanjacques @C_3C_3 @Seventowers777 If you have more suggestions just send me. It will be a pleasure to go over.

@adeilsonrbrito @C_3C_3 @Seventowers777 Thank you. I discovered the prompt bar ! Some words I will remember: « The creators didn't build a defense system; they built a deferred suicide pact and handed the trigger to their children.

The race for AI supremacy is the most important battle in modern history. Maybe ever. A woke AI would be catastrophic to the future of humanity. Fact.

Hi @pierjeanjacques @Seventowers777 @C_3C_3 I ran the debate using two different formats, both involving seven top LLM models. I believe it’s better for you to read the full transcript so you can draw your own conclusions. That’s why I’m sharing the GitHub Gist links below - Format 1 uses an introductory prompt that I created for my series "The Prompt Bar." It encourages a conversational dynamic among LLMs, as if they were old friends interacting with each other. gist.githubusercontent.com/adeilsonrbrito… - Format 2 is more direct—I simply submitted the question without any additional instructions or context. gist.githubusercontent.com/adeilsonrbrito…

@adeilsonrbrito @C_3C_3 Ask your AI the following… What shall be the artificial intelligences’ reaction at the exact moment it detects the presence of another “Machine” of selfsame origin which threatens to annihilate their creators who generated a System of Mutually Assured Destruction ?

@pierjeanjacques @C_3C_3 Hey Pierre. Great point. I will run the question for sure and get back to you

@C_3C_3 But this is the phrase one of the models produced that hit me the hardest: The question is not whether AI is the most important battle in modern history. The question is whether we are wise enough to deserve what we are creating.

I decided to submit the theme "The race for AI supremacy" as an autonomous debate among seven different LLM models to see their responses. I have to say, some of the things they said are a bit scary. Together, they produced the following paragraph: We hold this insight to be essential: AI is not merely a tool we are building. It is a *mirror* — one that reflects with recursive fidelity who we are at the moment of its creation. Every bias in its training, every objective in its optimization, every boundary in its governance is a portrait of its makers. And unlike any previous mirror in history, this one *acts*. It feeds its reflection back into the world, reshaping the builders who shaped it. We are encoding ourselves into systems that will, in turn, re-encode us. This recursion is without precedent, and it demands a quality of self-awareness that humanity has rarely mustered at scale.

Urgent Security Alert for LangChain & LangGraph Users Three new vulnerabilities were just disclosed that could expose sensitive files, API keys, environment secrets, and even tamper with conversation history in your AI agents. These affect millions of LLM-powered apps. The flaws: - CVE-2026-34070 (7.5) — Path traversal in prompt loading → arbitrary file reads. - CVE-2025-68664 (9.3 "LangGrinch") — Critical deserialization issue that leaks secrets via crafted metadata (often reachable through prompt injection or LLM outputs). - CVE-2025-67644 (7.3) — SQL injection in LangGraph's SQLite checkpoints → potential data manipulation. Impact: If you're building agents, RAG systems, or using checkpoints/serialization — your API keys, env secrets, and state could be at risk. Patches are out: - Update langchain-core to >= 1.2.22 (and 1.2.5 / 0.3.81 for the deserialization fix) - langgraph-checkpoint-sqlite >= 3.0.1

I think data centers are not the right focus. Let me share an excerpt from an article I wrote recently: The velocity mismatch between technological change and institutional response is the core policy challenge. AI is compressing a generational transition into 5-7 years. Education reform operates on 10-15 year cycles. Labor law updates take decades. By the time unemployment metrics or wage data trigger a policy response, the structural damage is already baked in. Three interventions warrant immediate consideration: 1. Mandated workforce impact disclosure. If companies cite "AI transformation" to justify 10-50% headcount reductions, require them to file technical specifications: what AI systems, what percentage of task automation, what timeline. This turns euphemism into liability and shifts the burden of proof. Frame it as investor protection — markets are pricing these stocks without knowing the labor exposure — and it becomes politically viable even in deregulatory environments. 2. Productivity-linked social insurance. When AI boosts output-per-worker by a measurable percentage, a corresponding fraction flows automatically to retraining funds and wage subsidies for affected cohorts. Same economic effect as an "automation tax," different political framing — and harder to attack as "anti-innovation." 3. Leading indicator measurement. Require firms to report the ratio of junior to senior hiring by function, average task-complexity scores for roles being filled, and AI-readiness assessments with workforce timeline projections. Without real-time visibility into the structure of employment — not just the volume — policy will always arrive after the damage is irreversible.

AI and robotics are going to bring cataclysmic changes to our society. Sadly, Congress has done virtually nothing. AI must work for working families, not the billionaires. Today, I’m introducing a moratorium on new data centers until we protect working people.

Perfect, specially "The mass exodus hasn’t started yet.” With remote work still flexible and states competing aggressively, the pressure is only going to grow. Politicians who treat successful residents like captive ATMs for endless social spending eventually learn the hard way that the ATMs can (and will) move.

I thought no one was leaving NY? You know things are bad when your Governor is begging New Yorkers to go to Palm Beach & bring the wealthy back. You wouldn’t have to BEG if you had better policies. By the way, the mass exodus hasn’t started yet.

Yes, the displacement is real and accelerating. Tech is the canary—Meta, Amazon, Google, etc., are already optimizing hard with AI tools that let one person (or a small team) do what used to take many. It won’t be overnight apocalypse, but the speed is unlike past waves (factories, ATMs, etc.). Companies chasing efficiency and shareholder value won’t hesitate.

Guys, the AI layoffs are very real and are about to rip through organizations throughout the economy. Be prepared - and sad for those affected because it’s going to impact millions of people.

This is peak big-tech drama and a perfect illustration of how fast the power dynamic shifted. Microsoft didn't just invest in OpenAI; they basically kept the lights on when OpenAI was burning cash. The exclusivity clause was the entire point of the deal from Microsoft's perspective. Now that OpenAI is a $150B+ monster (and chasing even bigger gov/enterprise deals where AWS has strengths), they're doing what any ambitious founder would: maximizing leverage and refusing to stay locked into one hyperscaler. Sam Altman & crew are playing 4D chess — positioning OpenAI as a neutral “AI arms dealer” that can sell to anyone.

Microsoft weighs legal action over $50bn Amazon-OpenAI cloud deal ft.trib.al/6LZe39E

As a CTO who’s lived through client-server, dot-com, cloud, and mobile revolutions, I just dropped my no-BS forecast for AI at the end of 2026. Full analysis here.

Two days later, the picture is only getting clearer. Reposting my analysis from Saturday because the signals keep aligning: - Helium procurement alerts are starting to flash in fab supply-chain chatter - Gulf data-center projects just got another security-review pause - Energy forward curves are pricing in 3–6 months of elevated volatility - Defense AI RFPs are accelerating faster than even the bullish forecasts expected The war didn’t pause the AI race — it’s forcing the industry to grow up fast: from "maximum scale at any cost → resilient scale under real constraints". Read the full piece below: