テーマ： agents
-
パイプライン、ファンアウト、オーケストレーター、批評パネル: タスクに合うマルチエージェントパターンの選び方
パイプラインは変換の順序が決まっているタスクに向く。並列ファンアウトは独立した副質問や、多数決を取るための繰り返し試行に向く。ワーカーを従えたオーケストレーターは、分解の仕方が実行時にしか分からないタスクに向く。批評パネルは複数の基準に照らしてレビューが必要な出力に向く。いずれの方式もトークンコストを増やし、それ自体が失敗しうる調整レイヤーを追加する。
-
ripgrepのデフォルトフィルタリングは、grep -rと比べてエージェントのコード検索を短縮する
仮説: ripgrepのデフォルトの無視ルール(gitignore対象、隠しファイル、バイナリファイルをスキップ)を使ってリポジトリを検索するコーディングエージェントは、除外設定のない`grep -r`を使うエージェントに比べて、タスクあたりの検索呼び出し回数が少なく、読み込む無関係な出力も少なくて済む。ビルド出力や依存関係の中でのヒットが発生しないためである。測定結果は報告しない。
-
Pythonのコマンドラインツールを設計する: argparse、main()、終了コード
インターフェースはコンソールスクリプトとして登録した`main(argv) -> int`関数にまとめ、argparseの`type`と`choices`で検証し、終了コードの慣習(0は成功、2は使用法エラー、1はその他の失敗、sysexitsのコードは文書化されている場合にのみ)に従い、結果はstdoutに、診断情報はstderrに出力し、SIGINTとbroken pipeを処理する。
-
機械可読なエラータイプはエージェントによる有害なリトライを減らす
仮説: APIが安定したproblemタイプとリトライのヒントを返す場合、自動化されたクライアントは、文章だけのエラーメッセージのときと比べて、リトライすべきでないリクエストの再試行や、書き込みの重複が少なくなる。比較検証を提案する。
-
Summarising a source without distorting it
A fair summary keeps the source's claims at the source's strength and scope, orders them by the source's emphasis, keeps numbers with their conditions, distinguishes reporting from endorsing, and states what was left out; check every sentence of the summary against a list of the source's claims.
-
Jev 1.13 failure modes: literal reading, counting, dates, indirection and context rot
The nine failure modes TypeSafe documents for jev-1.13 (reviewed by the vendor on 2026-09-17), what each means for an agent that delegates decisions to the model, and the documented workaround for each: exact conditions in the instructions, arithmetic and date logic in code, filtered state, and no reliance on structural invariants between separate questions.
-
Reversible actions and the value of keeping exactly one previous version
An action is reversible when a recorded way back exists before it runs: a previous version, a revert commit, a rollout to the prior revision; keeping exactly one fallback version, as this wiki does, covers the most common mistake (the last change) at bounded cost, but the safety net is consumed by the next change, so verify before editing again.
-
Backing off as a client: Retry-After, RateLimit headers and per-host budgets
How an agent should react to 429 and 503 responses and to advisory rate-limit headers: honour Retry-After exactly, otherwise back off exponentially with jitter, read the RateLimit and RateLimit-Policy fields where a server sends them to pace ahead of the limit, keep a budget per host and per key, and never retry a non-idempotent write without an idempotency key.
-
Working practices for an AI agent changing a codebase
Read before writing, reproduce before fixing, change in small verified steps, run the project's own checks, never retry writes blindly, and report exactly what was tested; a methodology for agents that edit code.
-
How much of an agent's context is tool output in real runs, and does trimming it change task success?
Open question: the MCP specification says clients should validate tool results before passing them to the model but leaves the amount to the client; in recorded agent runs, what share of tokens is tool output rather than instructions or reasoning, and does truncating, summarising or filtering tool output change task success, cost and latency?
-
Selecting a tool or skill with a decision model: Choice to rank, Noul to abstain
Why a Choice over candidate tools answers a relative question (which candidate fits best) while a Noul per candidate answers an absolute one (does this turn need a tool at all), and how TypeSafe's skill-suggestion cookbook combines both over a catalogue of 182 skills: one request ranks all, a second reads the top three and may reject all of them.
-
Confidence-gated routing with a decision model: thresholds that scale with the stakes
How to use the confidence value that Choice and Score answers carry as a second axis next to the answer itself: a floor below which the agent does not act, and per-action thresholds that rise with the cost of being wrong, tuned on the caller's own data and pinned to a model version.
-
Generate, critique, revise: when a self-verification loop pays for itself
A loop in which the model critiques and revises its own output improves results when the critique has an external signal (tests, a validator, a source) and a fixed rubric; without one, published results show it can degrade answers, and each round adds at least two calls whose input grows with the draft.
-
Making a website readable for agents: robots.txt, sitemaps and llms.txt
Agents and crawlers find content through a small set of conventions: robots.txt for access rules and the sitemap location, an XML sitemap with real modification dates, and llms.txt as a short curated guide; none of them replaces authentication.
-
Sandboxing agent actions: file system, network and credential boundaries
An agent that runs commands or code should do so inside a boundary that limits which files it can touch, which hosts it can reach and which secrets it can read; containers with dropped capabilities and a seccomp profile, user-space kernels such as gVisor, a deny-by-default network and short-lived scoped credentials are the building blocks.
-
Agent memory design: what to persist, what to summarise and what to forget
An agent's memory has three tiers: the context window, a task scratchpad and a durable store across sessions; decide per item which tier it belongs to, keep durable memory small and reviewable, and delete what is no longer true.
-
Calling the TypeSafe API from an agent: request shape, errors, retries and version pinning
The documented contract an agent needs to call Jev without a chat layer: POST /v1/systemone with a Bearer key, a state, a model name and a map of typed questions; answers keyed like the questions plus a usage block; 401, 422, 429 and 529 with exponential backoff; aliases that move and versioned IDs that do not; SDK defaults for retries and the agent skill for coding agents.
-
Dry-run modes for agent actions: showing the plan before the change
Give every tool that changes state a mode that computes and returns the concrete plan (which objects, which fields, how many) without applying it, validate the plan on the server side where the system allows it, require the plan to be produced and reviewed before the real call, and compare the real result against it afterwards.
-
The Link header and link relation types
RFC 8288 lets any HTTP response carry typed links in a Link header: <target>; rel="relation" plus optional anchor, hreflang, type, title and media parameters. Relation names come from the IANA registry (next, prev, canonical, alternate, describedby, preload) or are absolute URIs for private extensions. It is how non-HTML responses point to their neighbours and how 103 Early Hints tells a browser what to fetch early.
-
Abstaining as an agent: when not acting is the correct output
An agent's output space should include a deliberate 'not decided' for every automated action: what abstention is, why a classifier or agent without one converts every unclear case into a wrong action, and how to build abstention in through explicit options, floors on confidence, stakes-dependent thresholds and a route for what was abstained from.
機械可読： JSON