Category building · 8 artifacts
Agent Systems
Published work collected in the Agent Systems building.
Artifacts in this building8 published
PRThin skills, sharp tools — model-agnostic agent orchestrationThree thin skills plus two CLIs replace five Codex rulebooks; one bundle for every harness.ProjectSkillsCLI toolsPRCodex Orchestration SkillsI recently overhauled my Codex orchestration skills and cut the whole system back to just two: `teamwork` for bounded multi-agent work and `governed-delivery` for tracked repository changes that really need custody, exact-head review, or an external delivery gate.
This is a deliberate move away from the larger cast of named roles, runtime hooks, and state machinery I had built. Ordinary work should stay ordinary. Teamwork should load when parallel agents will materially help, and Governed Delivery should load when the risk of a merge, deploy, migration, or other external action earns the extra structure.
I am trying this setup now to see whether the smaller instruction surface makes Codex more predictable and easier to maintain without losing the boundaries I care about. The attached field note reflects the current in-progress v2 working tree; the released `main` branch is still v1.0.0, so this is an experiment rather than a claim that v2 has shipped.ProjectCodex SkillsTeamworkLECaching in LLM agent systemsI read [Luke Parker’s post](https://x.com/LukeParkerDev/status/2080045430109556772) and ended up looking more closely at a question I keep bumping into: if an orchestrator is long-running, does it keep getting prompt-cache benefits?
The distinction that helped me is that the thread, the harness context, and the provider cache are separate things. A dormant thread is not spending tokens or keeping a provider cache warm. When it wakes, the next request can still hit if the early prefix is unchanged and the provider’s current retention and routing rules line up, but compaction, changed tools, new early instructions, or elapsed time can make it cold.
I wanted one guide that keeps the model’s in-generation KV cache, cross-request prefix caching, semantic caching, and ordinary tool caches from blurring together.LearningLLM InferencePrompt CachingLEThe new rules of context engineering for Claude 5 models[Thariq Shihipar’s article](https://x.com/trq212/status/2080710971228918066) argues that newer Claude models change the context-engineering tradeoff: less duplicated rule-setting, more model judgement, interface design, progressive disclosure, and high-fidelity references.
One concrete reported result is that Anthropic removed over 80% of Claude Code’s system prompt for newer models with no measurable loss on its coding evaluations. The attached note keeps the article’s five “then/now” shifts together and maps the intended jobs of the system prompt, CLAUDE.md, skills and tools, references, and memory. It also keeps the evidence boundary visible: this is the author’s report, not an independently published benchmark.LearningContext EngineeringAgent ContextLETowards Automating Eval EngineeringI read [Viv’s “Towards Automating Eval Engineering”](https://x.com/Vtrivedy10/status/2079976006644072796), about a skill that builds evals from an agent repository and its traces.
What stands out is that it does not try to one-shot the eval. It interviews the user, builds a Harbor task, and then checks both the agent and verifier trajectories for shortcuts. I like the idea of turning real failures into durable tests; I just want the loop to stay small enough that the eval system does not become another thing to constantly maintain.LearningEval EngineeringAutomated EvaluationLEWhy We're BuzzingI read [jack’s “why we’re buzzing”](https://x.com/jack/status/2080056638820450400), about Buzz, an open-source workspace that keeps people, agents, conversations, code, and approvals in one signed record.
The part I keep thinking about is the context problem. Agents can only help with what they can see, and we lose a lot between chat, code, CI, and review tools. I am curious how much of Buzz requires replacing the whole workspace versus making the seams between the tools I already use more coherent.LearningAgent ContextCryptographic IdentityVIAgent Swarms — Model EconomicsMy dylanmccavitt/skills repo was relatively close to this proposed structure in the sense that routing agents correctly, was the focus for an autonomous loop. I''m taking pieces of this article and implementing it in to my skills repo with some missing gaps that this currently brought up.Visualmodel economicsAI agentsNOCleaning my agent configurationsA reflection on simplifying Codex agent configuration to reduce hidden context, improve predictability, and measure whether customization is actually helping.NoteCodexskills