AI agents

Overview

AI agents are systems that combine foundation models with tools, memory, planning, and execution loops to complete multi-step tasks autonomously or semi-autonomously. Unlike a single prompt-response interaction, an agent can decide what to do next, call external systems, use files or APIs, maintain context over time, and iterate toward an outcome. In practice, the newsletter references both core agent architectures and more applied "agentic workflows," especially in coding, research, operations, and internet-facing automation.

For AI Product Managers, agents matter because they change the unit of product design from isolated features to goal-driven systems. Instead of specifying every rule and orchestration path up front, PMs increasingly define objectives, constraints, tools, specs, and evals for systems that can adapt during execution. This shifts product work toward designing environments, permissions, feedback loops, and success metrics for autonomous behavior—while also introducing new concerns around safety, reliability, token spend, persistent state, and user trust.

Key Developments

• 2026-02-01: Andrej Karpathy warned that large-scale autonomous LLM agent networks connected through a global scratchpad create major security and coordination challenges, highlighting the risks of multi-agent systems at scale.
• 2026-03-11: Santiago argued that AI agents reduce the need for hand-coded orchestration and decision logic, making complex workflows faster for PMs to build and manage.
• 2026-03-17: Peter Yang said PMs must write specs for AI agents directly, rapidly build AI fluency, and rethink traditional software planning assumptions such as waterfall methodologies and even budget structures like token spend versus salary spend.
• 2026-03-27: Guillermo Rauch noted that coding agents perform best when they can install, run, debug, and deploy code freely, but require persistent compute to preserve state across sessions.
• 2026-03-29: Russell J. Kaplan at Cognition observed that AI agents are beginning to autonomously kick off tasks, signaling a shift from reactive tooling to proactive engineering systems.
• 2026-03-29: Peter Yang, echoing Karrisaarinen of Linear, emphasized that when teams can spin up many agents in parallel, clear product vision and user focus become even more important to prevent wasted execution.
• 2026-04-10: Philipp Schmid shared five principles explaining why senior engineers struggle with AI agents: treat text as state, hand over control, view errors as inputs, move from unit tests to evals, and design evolving agents instead of static APIs.
• 2026-04-19: Hugging Face was described as a go-to platform for AI agents because access to a large ecosystem of Spaces and specialized models expands what agents can build, run, and connect to.
• 2026-05-14: Greg Isenberg argued that AI agents are becoming primary buyers on the internet, making MCP servers strategically important for businesses that want to be discoverable and actionable by agents.
• 2026-05-25: Peter Yang described onboarding AI agents before hiring humans in order to learn role pain points faster and improve training loops over time.
• 2026-06-17: Philipp Schmid shared a free 5-day course on building AI agents, covering architecture, tool use, planning, memory, and evaluation with hands-on code and notebooks—evidence that agent building is maturing into a practical product and engineering discipline.

Relevance to AI PMs

1. Agents require PMs to design outcomes, not just screens or flows. PMs need to define goals, permissions, constraints, tool access, failure handling, and escalation paths. A strong agent product spec should state what the agent is allowed to do, what success looks like, when it should ask for confirmation, and how it should recover from ambiguity or errors.

2. Evaluation becomes more important than deterministic requirements.
Because agents operate in open-ended environments, traditional unit-test thinking is often insufficient. PMs should work with engineering on evals that measure task completion, tool correctness, safety, latency, cost, and user satisfaction across realistic scenarios.

3. Infrastructure and operating model decisions become product decisions.
Choices around persistent compute, memory, file access, MCP servers, static APIs, sandboxing, and token spend directly shape what agents can do. PMs should treat these as strategic product levers, not just implementation details, especially for coding agents, internal copilots, and workflow automation.

Related

• Philipp Schmid / Phil Schmid: Frequently referenced for practical guidance on agent principles, architectures, memory, tools, and evaluation.
• Evals: Central to agent quality because success depends on behavior across multi-step tasks, not just single outputs.
• Static APIs: Often contrasted with agents; agent-based systems are framed as evolving and adaptive rather than rigidly predefined.
• Cognition / Russell J. Kaplan / Devin: Connected to proactive and autonomous engineering agents that can initiate and execute work.
• Persistent compute: Important for agents that need long-running state, especially coding and operational agents.
• Specs: PMs are increasingly expected to write specs for agents, including goals, tool boundaries, and escalation rules.
• Token spend: A key operating constraint for agentic systems, especially when tasks involve long horizons, tool loops, or many parallel agents.
• MCP / MCP servers: Increasingly important for making products accessible to agents as discovery and transaction interfaces shift toward machine clients.
• Claude Code, Codex, Cursor, OpenClaw: Examples of agentic coding workflows and development environments.
• Anthropic, Gemini, Perplexity, xAI: Model and product ecosystem players shaping agent capabilities and interfaces.
• LlamaIndex, Harrison Chase, Hugging Face: Ecosystem enablers for retrieval, orchestration, tooling, and deployment of agents.
• Files / sandbox-at-vercel / Stripe / HubSpot / Salesforce: Examples of the external systems and environments agents may need to read from, act on, or integrate with.
• Waterfall methodologies: Referenced as poorly matched to the faster iteration loops required for agent-based product development.
• Reasoning framework / cognition: Related to planning, decision-making, and the mental-model shift behind agent design.
• Agent-first startups / SDR: Examples of business categories where agentic workflows are becoming core product strategy.