Notes on agents, software, and work.
Systems / October 17, 2026 / 10 min read
Adaptation of Agentic AI: where the change lives and whether it lasts
Agent adaptation spans post-training, memory, skills, and fixed-model harness changes, with practical tests for persistence, controls, and rollback.
Systems / October 16, 2026 / 9 min read
Agentic AI with LangGraph: build around state, control, and recovery
LangGraph earns its complexity for explicit state, branching, recovery, intervention, and durability; basic tool use often needs only a plain loop.
Systems / October 15, 2026 / 9 min read
Semantic Kernel multi-agent orchestration: choose by control contract
Semantic Kernel orchestration differs in routing, approval, history handling, and recovery; these control boundaries determine what fits your workflow.
Systems / October 14, 2026 / 8 min read
AWS multi-agent orchestration: choose by workload and control ownership
Choose an AWS multi-agent orchestration path by workload independence, orchestration ownership, deterministic control, observability, and cost.
Systems / October 13, 2026 / 10 min read
Databricks agentic AI: a production boundary scorecard
Databricks covers data, tools, evaluation, observability, policy, and state, and this scorecard maps the scope of each production boundary.
Systems / October 12, 2026 / 9 min read
Goose agent harness: a control-surface review and adoption scorecard
Goose's fit depends on its control surface: context loading, persistent state, tools, verification, permissions, and isolation.
Systems / October 11, 2026 / 10 min read
Latent collaboration in multi agent systems: the production decision behind the benchmark
One latent compression method cut communication cost 79.8% to 89.4%; production still depends on replay, evaluation, controls, and compatibility.
Systems / October 10, 2026 / 9 min read
Pydantic AI Harness: what is documented, what is proposed, and how to decide
Pydantic AI already has a light harness; the separate Harness package, open issues, and production-readiness tests answer different adoption questions.
Systems / October 9, 2026 / 8 min read
Strands vs LangGraph: choose by control, state, and recovery
Strands, LangGraph, or both fit different control boundaries across execution, durable state, recovery, observability, and deployment.