The Art of Building for Agents | S3E1

Instinct's agent-to-agent coordination is the most interesting thing a consumer AI app has shipped recently, and Michael and Jhanvi get into how they've been using it. Instinct lives entirely in iMessage. Michael can ask his Instinct to coordinate something with Jhanvi's, and the two agents handle the back-and-forth while each person approves what gets shared. They compare it with Muse, from data connectivity to the experience of interacting with the agent. They also debate where the moat sits when a feature like Instinct's is easy to copy and network effects can be bridged by integration, with accumulated memory about the user as a more durable advantage than any single feature. Finally, they cover the gaps that keep either app from replacing a daily driver, including limited calendar support and no way to read what your agent said on your behalf. The conversation then turns to what consumer AI means for the firms that buy enterprise tools. As people get used to simple, connected assistants at home, the pressure on enterprise AI policies around email and calendar access will grow. They close with model news: a new generation of cheaper, faster releases, the case for small fast models as top-of-funnel classifiers across large volumes of news and filings, and why token speed matters more as agents carry more context into every session.

Opus 5.5 shipped at 40% below Fable. Meta Muse explodes in popularity. It's a busy time in the world of AI.

OpenAI and Anthropic are calling for a slowdown in frontier AI development. Nvidia and Meta are not. The line runs between companies whose revenue depends entirely on frontier research and companies with tested businesses underneath them. Jhanvi and Michael take apart the incentives behind the debate, and they cover everything from the moat heavier regulation would build for larger players to the data center fights in states like Michigan, where towns are debating whether these centers do more harm than good. Then the question every firm rolling out agents hits in week one: who is the agent logged in as? Authenticating as the user keeps accountability clean, until that person leaves and the job breaks. So is the answer service accounts with pre-configured permissions? Plus the new favorite way of talking to agents, which surprises no one: email.

Data engineering in asset management was always a context problem, not a technology problem. The analyst knows what the data means, the data scientist knows how to query it, and the data engineer sits two steps removed from the domain, building pipelines against a business problem nobody translated cleanly. Michael and Jhanvi make the case that AI collapses that chain: you can now apply context at the last mile, at inference time, instead of spending the first 99 miles encoding it into a warehouse that turns into cement. The security master edge cases, the identifier mapping across 300 names, the non-GAAP KPI alignment against Visible Alpha, all of it fits in a cheap model's context window. So what does a data engineer do now? Build the context layer. Skills, agent instructions, tools, memory, and an observability layer that tells you which agents are failing and why before your users do, then harden pieces into deterministic code only once you actually understand them. Jhanvi and Michael also get into where this breaks: an agent harness that can't respect your data permissions, MCP connectors that authenticate as a service account instead of on behalf of the user, and the question of what accountability looks like when the work is done under an agent's name instead of yours.

Hedgineer's annualized AI bill has grown past $1M, which brought back up a long-running internal discussion at the company: how do you measure ROI on a token? Michael and Jhanvi work through both layers of the problem. The first is instilling good AI hygiene practices across the company — encouraging thoughtful model selection, session discipline, keeping irrelevant context out of the window to optimize on cost. The second is harder. Tag every session to the feature, product, or service it produced, and the company starts to look like a portfolio of assets, each with a cash flow you can discount, price, and reallocate capital against. Running a business through the lens of AI turns every operator into an asset manager. The complication is that building a product has never been cheaper and building a good product has never been harder. Everyone can prototype now, so the constraint moved from execution to taste, and killing good ideas in favor of great ones grows more challenging. They also get into why an observability layer and a skill library are worth more together than either is alone, and why some AI products are immediately understood and then never used. About Hedgineer Hedgineer is building the AI platform for institutional investing — deploying agents, skills, and data connectors directly inside hedge funds and asset managers to transform investment and operational workflows. The Hedgineer Podcast follows CEO Michael Watson and COO Jhanvi Virani as they navigate the frontier of AI adoption in finance, sharing unfiltered perspectives from the teams, guests, and problems they work with every day. Subscribe for weekly analysis on AI infrastructure and institutional finance. Watch the full episodes on Spotify at https://isht.ink/dFj5oaqbe or YouTube at youtube.com/@hedgineer. Audio available wherever you get your podcasts. Connect with us on LinkedIn at linkedin.com/company/hedgineer-io or reach out at [email protected]. Hedgineer.io