AI Agents
Everything related to agentic flows, harnesses, development with agents, building agents.
Build a bounded model failover policy that separates transport errors from refusals, preserves contracts, and checks context before resuming agent runs.
Prevent agent retries from repeating charges, messages, or creates by tying one durable idempotency key to each intended operation.
How to set deactivation thresholds, isolate control, stop agent work, preserve evidence, and define safe redeployment.
How to bind tenant identity to storage, caches, tool checks, runtime state, and quotas without mistaking logical partitioning for isolation.
Prevent agents from overwriting newer state by using conditional writes, recomputing after conflicts, and isolating replayed work from side effects.
Arrange repeated agent prompts for prefix reuse, avoid early cache-breaking edits, and measure cached tokens against the full input.
Apply an allowlist at trace collection, strip sensitive payloads and baggage, and retain only the fields needed to operate agent runs.
Cap nested agent retries by attempts, time, and side effects; retry only safe transient failures, then return a typed error or escalate.
Use overlap, call-time secret resolution, and usage evidence to rotate agent credentials without failing in-flight runs.
Contain browser agents with preflight allowlists, untrusted-content handling, scoped credentials, state checks, and last-step human confirmation.
A task-by-task procedure for isolating agent-requested code, limiting denial-of-service paths, brokering access, and testing failure.
Persist delayed work, define missed-run behavior, and make scheduled agent handlers reject duplicate delivery.
Separate streamed text from committed agent state, route tool events independently, and make disconnects end in an explicit error or resumable turn.
Choose by consumer: schemas for program-read results, free text for people, and a hybrid when both need the same agent output.
Use one agent for context-heavy work; delegate checkable search and summaries when exploration is large, then cap parallelism to your rate budget.
How head and tail sampling behave at agent-scale trace volume, what tail buffering costs, and how to keep decisions consistent across services.
Build a result boundary that rejects malformed, oversized, or ambiguous tool output before a model can read it.
Build reproducible prompt artifacts, log prompt and model versions, evaluate changes, canary releases, and roll back without redeploying.
Authenticate, deduplicate, persist, and queue webhook events before an agent runs, keeping sender responses fast and retries under your control.
When a while loop around a model call beats an agent framework, what frameworks actually supply, and the three requirements that flip the answer.
Three kinds of agent state — history, working state, durable knowledge — and where each belongs, how it is evicted, poisoned, read, and resumed.
How to have an agent provision a Cloudinary environment mid-session with one npx command, store the credential in a file, and claim it before it expires.
How to record an agent run — traces, tool arguments and results, prompts as sent — so any task can be reconstructed later without re-running the model.
How to decide what an agent may do unattended: reversibility, tool allowlists, per-session scope, short-lived credentials, and a human escape hatch.