Eric Siegel

Excited to see what Gianni Giacomelli and the kAIgentic team bring to market. Their model is a deliberate attempt to close one of the biggest gaps in the AI ecosystem today: Old-school executives — deep domain knowledge, operational expertise, and industry trust, paired with AI innovators — cutting-edge technical capability but limited exposure to real enterprise complexity. Combine these groups and you get a Collective Intelligence engine. Add AI as the processing layer, and it becomes Augmented Collective Intelligence (ACI) — a system where humans, machines, networks, and data work together to create value that neither side can achieve alone. kAIgentic is a textbook example of this shift. I’m excited to watch their trajectory.

3

Is more context always better for AI agents? A new paper on arXiv suggests the opposite. The study “Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents?” tests a common assumption in AI-assisted development: that adding repository instructions (like AGENTS.md or CLAUDE.md) helps coding agents perform better. The results were surprising: - More context didn’t improve performance Across multiple models and coding agents, adding context files often reduced task success rates compared to giving the agent no repository context at all. - Costs increased significantly. Those extra instructions also raised inference costs by over 20%, meaning teams paid more while solving fewer tasks. - Agents followed instructions too literally Context files pushed agents to explore more files and run more tests, but the added requirements often made tasks unnecessarily complex. - The key insight: less is more The researchers conclude that if context files are used, they should contain minimal, targeted instructions, not long lists of rules. Why this matters for the community and businesses building with AI: • More context ≠ better agents • Prompt engineering at the repo level is still immature • Efficiency and simplicity often outperform complexity As agentic development grows, this is an important reminder: the challenge isn’t just building smarter agents — it’s designing the right environment for them to operate in. Sometimes the best instruction is fewer instructions. 😉

25

3 Comments

Most search tools help surface standalone image files with metadata. But critical images inside PDFs, Word documents, PowerPoint presentations, and other document formats — be it charts showing quarterly performance, product diagrams in technical manuals, architectural blueprints in project docs — are all inaccessible through conventional search. https://lnkd.in/e57Seg2C

3

AI is accelerating "vibe-coding", but Margaret-Anne Storey warns of a hidden cost: cognitive debt. When we let agents build without oversight, we lose the mental model of our own systems. Stop the fragmentation of shared theory. Dive into the risks at https://buff.ly/eKFDkj5 #ProductManagement #AI

7

An AI policy should map to controls. Otherwise it's mostly a statement of intent. The policy says: Do not upload sensitive data. The control asks: Which tools can access sensitive data, and how is that enforced? The policy says: Human review is required. The control asks: Who reviews, what do they see, and can they override? The policy says: AI-generated content must be labeled. The control asks: Where is the label applied, how durable is it, and what evidence is kept? The policy says: High-risk decisions require oversight. The control asks: What counts as high-risk, what is logged, and who owns the outcome? This is the governance gap. A policy tells people what should happen. A control changes what can happen. A lot of companies have the first. Fewer have the second. For production AI, every policy statement should have an operating counterpart: rule owner system control evidence exception path monitoring review cadence If a policy can't be mapped to a control, it may still be useful guidance. But it's not production governance yet. --- Which AI policy is hardest to translate into a real control? ♻️ Repost this to help someone building, funding, or scaling AI systems. Follow Guillaume for practical thinking on AI, data platforms, production systems, and better business decisions.

74

76 Comments

Most AI systems today optimize. But they do not decide. That’s the structural problem. After months of research, modeling, and system design, I’ve published a new framework: Decision Infrastructure A governance-first mathematical layer for AI systems operating in high-stakes environments. 🔗 https://lnkd.in/d6EMDyVG This work introduces: • A unified decision equation integrating probability, impact, irreversibility, and uncertainty • A governance-constrained decision model (G(d)) • A human-final authority layer (H(d)) • A dynamic, time-aware risk system The key insight: AI failure is not a prediction problem. It is a decision authority problem. We are moving from: Optimization → Governance Prediction → Decision Authority Automation → Controlled Execution This framework is designed for: • Financial systems • Energy infrastructure • AI governance • High-risk decision environments The future of AI will not be defined by how well systems predict. But by how responsibly they decide. Feedback, critique, and collaboration are welcome. It is a governance-first Decision Authority Infrastructure.

3