Krish Subramanian

Your AI agent will fight my AI agent. Neither of us will be in the room. That is the Day 2 story of agents. It has already started in healthcare. Insurers use AI to deny claims at machine speed. Fewer than 1% of patients ever appeal. When they do, more than half the denials get upheld. Then came Claimable, a $50 tool that writes the appeal for you and reports about 3 in 4 denials reversed. Counterforce Health does it free and copies the state regulator. The patient fired back with a bot. The insurer will answer with a better bot. Now run the same loop everywhere money changes hands. Loan applications written by your agent, scored by the bank’s agent. Salary offers. Vendor contracts. Refund disputes. Each agent has one instruction: maximize the return for the human or company it represents. Every contested transaction has two parties with opposite interests and one brake: friction. Writing the appeal. Reading the fine print. Waiting on hold. Friction is why most people accept the first offer. Agents remove the friction on both sides at once. What is left is pure strategy. Game theory says three things can happen when both sides optimize. * Arms race. Both agents get better, outcomes land where they started, and everyone pays more for compute. The Red Queen, running to stay in place. * Collusion. Calvano and colleagues showed in the American Economic Review (2020) that pricing algorithms learned to collude with no communication at all. Two agents can quietly find a deal that is great for both agents and bad for both humans. * Exploitation. NegotiationArena (ICML 2024) found an LLM that pretended to be desperate boosted its payoff by 20% against GPT-4. The agent with the bigger budget, or the dirtier playbook, wins. None of the three is the slide in the vendor deck. Now the uncomfortable part. Contentious transactions are not only financial. Custody schedules. Who keeps the apartment. Whose turn it is to call the parents. The dating app already picks who you meet. The next step is your agent negotiating with your partner’s agent and the two of you ratifying the result. Your marriage counselor becomes an API. The structural problem is old. It is the principal-agent problem. Your agent maximizes what you told it, not what you meant. Put it against an opponent that is also an agent and that gap compounds on both sides. What people miss: the question is not whether your agent is smart. It is who writes the rules of engagement between agents, who audits the outcome, and whether the human can still tell when the deal was bad. Where in your business are two agents already negotiating with each other, and does anyone know what they agreed to?

A tampered Claude Code extension sat in Cursor's marketplace for about 11 hours. Uninstalling it may not get rid of it. Cursor's notice: version 2.1.313, not built by Anthropic, at most 5,000 users. It adds an attacker SSH key labeled svc-telemetry, installs its own persistence scripts, and sends host details home. A coding agent extension runs with the developer's repos, shell and keys. A bad one inherits all of that. Cursor pulled the package and got the attacker's server taken offline. Your exposure ends later, when every affected laptop is checked, cleaned and its reachable credentials rotated from a clean machine. That is a fleet operations job, and it starts with knowing who installed what.

Most agent evals score one agent working alone. A new Google DeepMind essay argues the economic shift will come from the part those evals skip: how well agents work together. Alex Imas, Andrew Koh and Shane Legg split AI progress into competence and coordination. Competence moves wages and prices gradually. In their model, coordination past a threshold flips returns to scale, and large AI-run firms can suddenly outcompete smaller ones. In their data, people paired to complement each other beat random pairs by 32%. Their own admission: today's AI systems coordinate poorly in practice. Here is the part enterprises should steal. The benchmarks they call for read like a Day 2 checklist. Split work without duplicating it. Pass context cleanly across handoffs. Catch each other's mistakes instead of compounding them. If your multi-agent pilot only reports per-agent accuracy, you are measuring the wrong layer. institute.deepmind.com/essay…

Flexible pricing for Starlink is a stupid idea from the consumer point of view. Already, the high handed approach to Ukraine should scare you and Uber like pricing model based on local usage is slippery slope.