Kubai Kevin

Older Posts — Page 5 of 19

Agent architecture: the new model trap

Newer models made prompt chaining obsolete but not policy enforcement. How to keep a deterministic layer around a capable planner.

17 min read · 27 September 2026

Tech debt: when half your stack is unfamiliar

A repeatable process for debugging, verifying and governing tools you didn't write, with baselines, contracts, environment measurement and tool cards.

14 min read · 26 September 2026

Portfolio projects hiring managers skip

Portfolio projects often fail screening for legibility, not code quality. Here is how to restructure a repo so a reviewer understands it fast.

19 min read · 25 September 2026

APM is blind to your AI agents

Traditional APM monitors transport, not semantics. See how to instrument retrieval quality, token cost, and answer correctness in AI agent pipelines.

12 min read · 25 September 2026

Multi-agent systems: when they quietly degrade

Multi-agent LLM systems drift into slow, costly behaviour without failing loudly. Here is what to instrument, what to budget, and how to catch the slide.

12 min read · 24 September 2026

World models won't replace your backend

World models add a high-rate prediction stream to your system. Treat it as a sensor: schema, retention, backpressure, and rollback, not a new stack.

15 min read · 24 September 2026

AI features: M-Pesa, Paystack, Flutterwave traps

Payment webhooks from M-Pesa, Paystack and Flutterwave disagree on timing, idempotency and error semantics. Here is how to normalize them before an AI feature reads them.

12 min read · 23 September 2026

Tool calls: why retries stampede

Why per-client backoff fails under fan-out tool use, and how shared token buckets, single-flight and one-layer retries fix it.

12 min read · 22 September 2026

Agent compliance: the layer you build after failure

Pre-execution policy gates miss most agent compliance failures. Build a decision trace that reconstructs what happened, why, and what would have stopped it.

16 min read · 22 September 2026

Sandboxed computer-use agents with gVisor

How to contain computer-use agents using gVisor, egress allowlists, and short-lived credentials, with code, failure modes, and planning figures.

14 min read · 21 September 2026

When your agent goes silent: degrade gracefully

How to keep agent-backed features useful when a model is slow or confidently wrong: deadlines, fallbacks, circuit breakers, validation, and how to measure each.

13 min read · 21 September 2026

CQRS: when it pays off and when it breaks

CQRS trades strong consistency for independent read and write scaling. Learn the failure modes, how to measure projection lag, and when a read replica is enough.

12 min read · 21 September 2026

Rust rewrites: where they pay off, where they don't

A practical framework for deciding whether a Rust backend rewrite is worth it, with code, failure modes, and measurement steps.

12 min read · 21 September 2026

Idempotency keys: what they fix, what they don't

How idempotency keys work, where they fail, and the failure modes they don't cover: concurrent retries, key reuse, expiry, and partial failures.

14 min read · 21 September 2026

Golden paths that don't handcuff

How to build internal scaffolds and CLIs that stay reversible: pinned versions, a compatibility contract, drift detection, and a tested eject path.

12 min read · 21 September 2026

Golden paths become load-bearing walls

Why thick internal SDKs for AI features become migration hazards, and how to design a thin runtime that enforces model ID, prompt version and embedding dimension.

13 min read · 21 September 2026

Offline agent workflows for last-mile ops

How to build an offline-first field agent client: IndexedDB schema, UUIDv7 identity, background sync, idempotent batch upload, and clock-drift handling.

15 min read · 21 September 2026

Dead-letter queues: stop poison messages

Design retry caps, DLQs, and alerts so a poison message is isolated in seconds. Includes working Redis Queue code and pytest tests.

15 min read · 21 September 2026

Agent compliance in African fintech: the CBN gap

How to map autonomous agent behavior onto CBN, NDPA and SEC obligations using a three-layer split: model, deterministic policy gate, append-only audit.

16 min read · 20 September 2026

Audit logs: compliance vs performance

How to design audit logging that survives crashes and preserves per-entity order without wrecking p99 latency, plus how to measure the overhead.

13 min read · 20 September 2026

AI Supply Chain Security: Trust Boundaries for Model Artifacts

How to reason about trust boundaries for AI artifacts when pickle, torch.load and remote code execution are the format, plus a practical detection stack.

16 min read · 19 September 2026

Computer-use agents: the permission trap

GUI agents inherit a human's full permissions and leave no semantic audit trail. Here is how a policy layer, a reduced interface and structured logs fix that.

15 min read · 19 September 2026

Agentic AI cost control: 7 tools, one backfire

How to control agentic AI spend: ground-truth billing data first, attribution second, and why LLM optimizers over billing data backfire.

15 min read · 19 September 2026

Scope agent permissions per task

Most agent permission models fail because credentials are scoped to the agent, not the task. Here is how to issue short-lived, task-scoped tokens and what to measure.

20 min read · 19 September 2026