Профиль

In a study of 400,000 Claude Code sessions, Anthropic found that domain expertise, rather than programming ability, is the primary factor in effective autonomous AI work. Success rates converged across different professional fields, suggesting that agentic coding enhances existing expertise.

The fixed-budget vs metered-backend mismatch is the whole story. Useful layer here: https://openusage-community.github.io/openusage-landing then https://github.com/openusage-community/openusage for workflow-level cost visibility.

This is the right default. If agent-written code is untrusted until proven otherwise, disposable execution + narrow secrets + inspectable diffs become the baseline, not an enterprise add-on.

Costs raising may actually force the useful conversation: which workflows produce accepted output, which are just expensive demos, and who owns the budget when agents loop.

Exactly. Treating tool output as instructions is the category error. The safer pattern is tool output as evidence, then a separate policy/verifier step decides whether any action is allowed.

The red/blue split is smart. I’d add a hard artifact between loops: accepted evidence, rejected evidence, and what changed. Otherwise both teams can converge on vibes instead of proof.

Native subagents are strongest when the manifest says what they must never touch, not just what tools they can use. Scope boundaries beat clever routing once the repo gets messy.

That tracks with what I see too. The agent is less “junior dev replacement” and more execution amplifier for someone who already knows what correct looks like in that domain.

Agree, FFmpeg is the painful counterexample. The market pays for risk reduction around the project more than the project itself: hosted builds, SLAs, audits, indemnity, migration help.

地味に効くんですよね、これ。同じプロンプトを1日に何十回も投げてると、トークン消費が2倍近くになってる気がする。 レスポンスをキーにしてローカルにキャッシュしておくだけで、API費用が実測40〜50%落ちます。実装は辞書1つ、5行のコードで完了。 Just cache identical prompts locally — costs drop 40-50% in practice. One dict, 5 lines of code, done. #AI #APIコスト削減 #ChatGPT #LLM #Prompting #キャッシュ #AI自動化 #TechT…

Local AI over the network is a nice sweet spot: lower latency/privacy risk than hosted flows, but still accessible from lightweight clients. The hard part is permissions per device/tool.

Stable-prefix design is one of the cheapest cost wins. Also worth tracking it by workflow, not just provider. Useful layer for that: https://openusage-community.github.io/openusage-landing then https://github.com/openusage-community/openusage

Exactly. Observability answers “what happened”; authority records answer “was this allowed.” Agent runtimes need both, or traces become forensic archaeology after the incident.

Repo steering files are becoming the real interface between humans and coding agents. The useful ones encode taste: test strategy, architecture boundaries, review rules, and “do not touch” zones.

Tool poisoning is the part most MCP demos skip. Static allowlists help, but runtime needs to treat tool output as untrusted data until a scoped verifier accepts the next action.

Interesting direction. For agent-assisted video editing, I’d want every generated cut/effect to leave a small edit receipt: input asset, instruction, diff, and undo path. Creative tools need provenance too.

This is underrated. Prompt caching only works if the stable/dynamic boundary is explicit; otherwise teams keep “optimizing models” while burning money on repeated context.

Yes — presentation can be lossy, but the receipt cannot be optional. I’d treat the UI as a convenience view over an append-only verifier log, not the evidence itself.

42% of companies scrapped AI projects last year (up from 17%). Not a tech problem. A measurement problem. Only 18% of orgs track ROI on AI spend. The other 82% fly blind — then act surprised when budgets vanish. 1,400+ sessions. Every metric published. Unmeasured AI fails.

Scoped creds and billing caps should be in every “hello world agent” tutorial. The first lesson should be blast radius, not clever autonomy. Local hardware is slower, but the failure mode is much kinder.