📰 Dev.to · Gabriel Anhaia

Agent decision quality decays quietly across steps and model updates. Build baselines, online sampling, and drift alarms that catch it early.

Handoffs vs a shared bus, message schemas, and the budget guard that stops two agents from talking each other into a bill.

An agent is not a web service. Package it, decide statelessness, cap concurrency, and scale a workload that blocks on the model.

Every tool call an LLM agent makes is a trust boundary. Least-privilege credentials, allow-lists, output validation, and sandboxing in real Python.

Prompt injection is unsolved. Meta's Rule of Two says pick at most two of three risk legs per agent session. Here is how to enforce it.

LangGraph, OpenAI Agents SDK, Claude Agent SDK, Microsoft agent-framework, CrewAI, Pydantic AI: a pick-by-need guide for 2026.

Typed dependencies, validated outputs, and tests that never touch the network. How Pydantic AI puts static types around a stochastic model.

Roles, conversations, or graphs. A field guide to the 2026 multi-agent frameworks, honest verdicts, and when one agent beats a crew.

Agents, handoffs, guardrails, sessions. The abstractions that help you ship, and the ones that hide behavior you need to see.

Fixed workflows, fully autonomous loops, and constrained autonomy. How to pick the least-powerful design that still solves the task.

Turn a goal into a plan of typed steps, learn why structured plans beat prose, and re-plan cleanly when a step fails.

Vague tool schemas make agents call the wrong function. Here is how descriptions, strict validation, and tool_choice fix misfires.

Build a real tool-calling agent from scratch: define tools, run the loop, feed results back, and stop cleanly. No framework.

Wrap a Go adapter with a fallback that serves stale cache when the downstream dies, and decide degrade-vs-fail per use case.

Keep HTTP status codes out of your Go domain. Typed sentinel errors, errors.As, and one translation function at the handler edge.

Exponential backoff with jitter in Go belongs at the adapter, not the domain. Respect context deadlines, retry the right errors, keep the port clean.

Wrap an outbound Go adapter with a circuit breaker. Closed, open, half-open states, and a use case that never learns the breaker exists.

Build an idempotency-key store in Go with first-write-wins and response replay, and put the check at the right layer of your handler.

Emit events in the same transaction as your state, then relay them with a Go poller. At-least-once delivery, no Kafka, no dual-write bug.

Import slog in your Go domain and you weld a logging library to your business rules. A narrow port keeps it at the edge, where it belongs.

Kill time.Now() calls buried in Go business logic. A Clock port, a real adapter, a fake, and tests that stop flaking on midnight.

google/wire builds your Go dependency graph at compile time. Provider sets, generated injectors, and ports wired to adapters with no reflection.

Spin up a real Postgres in your Go tests with testcontainers-go, get the lifecycle and readiness right, and keep the suite fast on every push.

Stamp version, commit, and build time into a Go binary with -ldflags -X, and read the git revision back from runtime/debug BuildInfo.