Notes from building agent systems
Evaluation harnesses, orchestration patterns, memory design and the failures behind them — written up while the details are still fresh.
The 8.3x Usage Gap Is an Integration Problem, Not an Adoption Problem
Licences measure procurement; usage measures integration. The gap between them is structural, and it lives in the tool surface between the agent and the systems of record.
Settlement Versus Authorisation: What Agentic Trading Reveals About Action Limits
Agents given cash to trade often buy and sell repeatedly: the failure is not the strategy but the absence of a settlement boundary between intent and recorded position.
Durable Execution Is Not Durable Reasoning: Checkpointing Agent State Without Losing Intent
Persisting an agent's step index is not persisting its task state. Systems that checkpoint execution but not intent resume into fluent, incoherent runs.
Durable Execution Is the Real Infrastructure Requirement for Long-Running Agents
Funding and product moves around durable execution point at the actual hard problem in agent infrastructure: deterministic replay and idempotent effects for runs that span hours and external side effects.
Legacy Estates Are an Agent Amplification Problem
Agents do not invent new vulnerabilities in legacy estates; they raise the throughput of exploitation. The fix is rate and scope limits at the agent boundary, not a rewrite.
Agent-Readable Interfaces: Designing Tool Surfaces for Machines Before Humans
When platforms open APIs to third-party agents, the interface contract must be designed for machine callers first, with explicit idempotency, quotas and error semantics.
Payment Authority for Agents: Scoped Spend Credentials Beat Stored Cards
Stored cards grant unbounded, reusable authority. Scoped spend tokens bind amount, merchant, expiry and task, making agent payments revocable and auditable.
State Management Beats Memory: A Design Note on Agent Persistence
Most agent "memory" failures are state-management failures. Retrieval over past turns cannot enforce invariants; a transaction log can.