12/01/2027
Agent Evals That Matter: Test the Failures a Demo Never Shows
Evaluate AI agents with verified outcomes, deterministic invariants, stateful failure scenarios, repeated trials, and cost per successful task—not polished prose.
[ Archive ]
Blog
Practical notes from current work and experiments across architecture, inference systems, and infrastructure.
05/01/2027
A Model-Agnostic Agent Harness Needs More Than an API Adapter
Design a model-agnostic agent harness with explicit provider capabilities, durable execution contracts, domain validation, safe fallbacks, and workflow-level evals.
29/12/2026
Live Islands in Static Reports: Real-Time UI Without Rewriting the Document
Add live islands to static reports with shared subscriptions, scoped data access, versioned updates, stale-state UX, and exports that preserve observation times.
22/12/2026
Model-Agnostic AI Artifacts: Give the Model a UI Contract, Not Your Frontend
Build model-agnostic AI artifacts with declarative schemas, trusted rendering, semantic validation, revisioned edits, and structured feedback to the model.
15/12/2026
Autoscaling AI Workers: Queue Length Is Not a Capacity Plan
Autoscale AI workers using queue age, work estimates, runtime headroom, and safe draining. Separate admission and placement from adding more replicas.
08/12/2026
Skills and Adaptive Tool Visibility: Discovery Is Not Permission
Design skill-based agents with progressive tool discovery, versioned procedures, stable tool profiles, and authorization that is independent of visibility.
01/12/2026
Agent Context Compaction: Treat Summaries as Checkpoints
Build safer agent context compaction with validated summaries, explicit history coverage, atomic checkpoint commits, and recovery that never drops history silently.
24/11/2026
On-Demand Subagents: Delegate Work, Not Authority
Design on-demand subagents with narrow tasks, scoped evidence, bounded budgets, cancellation ownership, and results the parent can verify before publishing.
17/11/2026
Browser Agents Are Identity Infrastructure, Not Just Automation
Build safer browser agents with scoped sessions, explicit ownership, human handoff, evidence versions, and recovery that does not casually reuse identity.
10/11/2026
Redis in Production: Stop Calling Everything a Cache
Choose Redis patterns by failure semantics: caches, queues, Pub/Sub, Streams, leases, eviction, and persistence each need a different recovery contract.
03/11/2026
AI Code Sandboxes: Docker Is a Starting Point, Not a Complete Security Story
Design AI code sandboxes around untrusted execution, network policy, resource limits, cleanup, and isolation—not just a container that starts successfully.
27/10/2026
Context Engineering for AI Agents: Build a Working View, Not a Bigger Transcript
Design agent context as a bounded working view with authoritative constraints, evidence references, output reserves, and valid tool-call relationships.
20/10/2026
Human-in-the-Loop AI: An Approval Button Is Not an Authorization System
Build human-in-the-loop AI approvals around exact proposals, resource versions, expiry, dispatch revalidation, and clear UX for uncertain external outcomes.
13/10/2026
Python RQ in Production: Simple Queues, Explicit Guarantees
Use Python RQ in production with durable task records, safe retries, JSON serialization, queue fairness, and a clear boundary between jobs and workflows.
06/10/2026
Prompt Caching for AI Agents: Optimize the Prefix, Not the Dashboard
Improve prompt caching with stable prefixes and honest measurement. Distinguish request hits, token reuse, provider accounting, and actual cost savings.
20/06/2026
Deterministic Infrastructure for AI Agents: Make Every Action Accountable
Design reliable AI agent infrastructure with explicit action contracts, ownership, idempotency, bounded execution, and honest recovery from unknown outcomes.