Hedgineer

Leveraging Agentic Memory Stores and Background Context Synthesis One of the hardest problems scaling enterprise AI agents is growing persistent context without bloating the token window over time. Anthropic's new memory store feature addresses this by acting as a targeted text database that logs specific agent interactions based on firm-defined rules, capturing user corrections, isolating proprietary formulas, flagging what's worth retaining. A background process called "dreaming" then periodically compresses those raw session logs, strips redundancy, and outputs a cleaner context layer the agent can use efficiently. Catch the full conversation tomorrow on Season 3, Episode 5 of The Hedgineer Podcast hosted by Michael Watson and Jhanvi Virani. https://lnkd.in/gNAtQdUF New episodes every week. Find us wherever you listen, and catch the video version on YouTube and Spotify. Hedgineer.io #AIEngineering #EnterpriseArchitecture #ContextOptimization

Architecting Model-Agnostic Platforms to Eliminate Enterprise Vendor Lock-In Vendor lock-in is a real concern, but it's more solvable than most teams think. The key is to avoid letting your custom logic live inside any vendor's infrastructure. If your workflows, skill libraries, and tool interfaces are hosted on a vendor's internal platforms, switching providers means starting over. That's where the riskiest exposure is. But if you keep your custom skill libraries in client-owned GitHub repos and build your tool interfaces against open MCP spec, your capabilities travel with you, not with the vendor, and pivoting to a competing model or an open-source alternative doesn't mean discarding years of engineering work. Catch the full conversation tomorrow on Season 3, Episode 5 of The Hedgineer Podcast hosted by Michael Watson and Jhanvi Virani. https://lnkd.in/gNAtQdUF New episodes every week. Find us wherever you listen, and catch the video version on YouTube and Spotify. Hedgineer.io #EnterpriseArchitecture #AIEngineering #ModelContextProtocol

We shipped Claude Code Auto Mode across our engineering team. Auto Mode lets Claude Code stop asking for permission on every tool call. A classifier evaluates each action against your policy and decides: run, ask, or block. The default is paranoid by design, great for solo work, not for a team shipping across multiple clients and cloud tenants. We designed a centralized policy, pushed it through managed settings, and measured the before vs. after with OpenTelemetry. - Auto-approval went from 82% to 95% without loosening a single rule - The classifier handled 21,000 more tool decisions - Engineers stopped creating their own "always allow" rules, those dropped 89% The biggest insight: the OTEL data was the most important input for designing the policy itself. Without it, you're guessing what to block. With it, you're reading what engineers already do and what the classifier doesn't yet trust. Full writeup with the policy template, telemetry breakdown, and what we're changing next at https://lnkd.in/g86zxz3G

AI Dev Days | Anthropic & Google Anthropic's developer day was almost entirely focused on enterprise infrastructure: agents, memory, and state management. Google’s I/O updates pointed toward consumer experiences and broad adoption. For technology and investment leaders choosing which ecosystem to build on, that divergence can be a signal about where each lab thinks enterprise value will actually accrue. Catch the full conversation tomorrow on Season 3, Episode 5 of The Hedgineer Podcast hosted by Michael Watson and Jhanvi Virani. https://lnkd.in/gNAtQdUF New episodes every week. Find us wherever you listen, and catch the video version on YouTube and Spotify. Hedgineer.io #FundamentalResearch #InvestmentManagement #AIEngineering

Season 3 Episode 5 Available NOW! Anthropic and Google both had major dev days recently, and their announcements say a lot about their business strategy. Anthropic is going deeper into enterprise: agents, persistent memory, and a new capability called agent dreaming. Google is going wide, embedding AI across every consumer surface it has. In this episode, we break down what each set of announcements actually signals about where the model race is headed. Listen to Season 3, Episode 5 now: https://lnkd.in/gNAtQdUF Hedgineer.io #AIEngineering #AssetManagement #ModelContextProtocol Michael Watson Jhanvi Virani

New Episode Tomorrow! Anthropic keeps shipping developer tools — MCP, skills, agentic memory — that aren't tied to their own models. Even Claude Code can run on any underlying model. This means that if you architect your company’s AI layer on these primitives correctly, you can mitigate dependencies on a model vendor. You're building portable infrastructure that happens to run on Claude today and can run on something else tomorrow. Catch the full conversation tomorrow on Season 3, Episode 5 of The Hedgineer Podcast hosted by Michael Watson and Jhanvi Virani. New episodes every week. Find us wherever you listen, and catch the video version on YouTube and Spotify. https://www.youtube.com/ ⁨@hedgineer⁩ Hedgineer.io #AIEngineering #EnterpriseArchitecture #ModelContextProtocol

 AI Accountability: Your Cultural Pillar for Client Business Scaling AI-native workflows requires a cultural pillar to match the technical one: the "Do You Stand By This" (DYSB) protocol, which holds that the human sender is the sole author and owner of any output, regardless of how much a model generated. In this clip, Jhanvi explains how building a culture of absolute accountability is what stands between a functioning agentic workflow and pure chaos. Catch the full conversation on Season 3, Episode 4 of The Hedgineer Podcast with co-hosts Michael Watson and Jhanvi Virani and special guest Mitchell Troyanovsky. New episodes every week. Find us wherever you listen, and catch the video version on YouTube and Spotify. https://lnkd.in/gu6JjEE2 Hedgineer.io #AIGovernance #ProfessionalServices #AssetManagement #AIStrategy

Recursive AI: Better Than Intelligence? Traditional RAG and append-only context management eventually break down under noise and state degradation. The "Better Intelligence" principle takes a different approach: using Recursive Language Models (RLMs) to let the agent decide what information stays in its own context. That recursion is what allows models to sustain performance over long-duration tasks without constant human recalibration. In this clip, Mitch from Basis explains why a model's ability to regulate its own state is the real unlock for continual learning in production. Catch the full conversation on Season 3, Episode 4 of The Hedgineer Podcast with co-hosts Michael Watson and Jhanvi Virani and special guest Mitchell Troyanovsky. New episodes every week. Find us wherever you listen, and catch the video version on YouTube and Spotify. https://lnkd.in/gu6JjEE2 Hedgineer.io #AIEngineering #LLMs #RecursiveAI #MachineLearning

Basis Platform: AI for Accountants Moving AI from a playground to production requires more than better models; it requires statefulness and governance artifacts baked in from the start. To automate an end-to-end process, a system has to capture the context, the integrations, and the audit trail at the same time. In this clip, Mitch from Basis explains why specialized platforms are replacing generic chat interfaces as the "system of record" that professionals actually need when a regulatory audit comes around. Watch the full conversation on Season 3, Episode 4 of The Hedgineer Podcast with co-hosts Michael Watson and Jhanvi Virani and special guest Mitchell Troyanovsky. New episodes every week. Find us wherever you listen, and catch the video version on YouTube and Spotify. https://lnkd.in/gu6JjEE2 Hedgineer.io #EnterpriseAI #Fintech #AuditCompliance #AIWorkflows