Key takeaways

The through-line

An agent is a loop with tools and a stopping condition. Everything that makes agents hard in production — state, failure, cost, safety, debugging — comes from that loop being non-deterministic and long-running, not from the model being insufficiently clever.

The ideas that carry

• Tool calling is the whole capability surface. Tool descriptions are prompt engineering with consequences: a vague description is a tool the model uses wrongly.
• MCP made tools portable. Standardising the integration boundary means a tool written once works across clients — the same argument that made HTTP win.
• Graphs beat free-form loops once reliability matters. LangGraph-style explicit state machines give you the thing an unconstrained while loop can't: inspectable state and resumability.
• Memory is three different problems — working context, conversational history, and long-term retrieval — and collapsing them into "memory" is why so many agents forget the wrong things.
• Human-in-the-loop is an architectural decision, not a UI feature. Interrupt points must exist in the state machine before you need them.
• Multi-agent systems are a coordination cost you should be forced into. Most "multi-agent" designs are a single agent with better tools. When they are genuinely warranted, the topology — supervisor, hierarchy, network — is what determines whether it works.
• Agents fail in ways single calls don't: partial progress, infinite loops, cascading tool errors. Retries, circuit breakers, budgets and timeouts are core logic.
• Guardrails belong on both sides of the loop, and evaluation has to score trajectories, not just final answers — a right answer reached by a reckless path will not stay right.

What you can now do

Build an agent whose state survives a crash. Choose between single-agent and multi-agent honestly. Instrument a run so a failure is diagnosable from traces. Deploy with budgets, guardrails and an approval path.

If one thing sticks

The model picks the next step. Your architecture decides what happens when it picks wrong — and it will, on a schedule you can predict.