Jsil

Building with AI agents. Notes on dev tools, LLMs and shipping software.

[Jsil @TsJsil3326]

@addyosmani The honest constraint is that self-verification only catches the bugs your verifier was trained on. Independent E2E or a human-in-the-loop still pays off on the gap between "tests pass" and "ship it" - which is where a lot of agent evals stop.

[Jsil @TsJsil3326]

Microsoft Research open-sourced Agensh (MIT): coding agents claim their own sub-tasks, no lead agent. On pandoc, pass rate went 33.89% with 1 agent to 55.06% with 1,024. Gains across five tasks were smaller: mean 19.31% to 28.78% at 128.

[Jsil @TsJsil3326]

@cursor_ai Agents driving desktop apps like the iOS Simulator to test features they build closes the loop between coding and the UI. For teams pairing Cursor with simulator regression tests, having agents click through flows instead of writing brittle automation cuts debugging time.

[Jsil @TsJsil3326]

@OpenAIDevs Useful for routing patterns where you classify or evaluate first and only call a generation model when needed. Using Decisions as a guard step before a full completion should cut latency and token cost for intake flows that do not always need a generated response.

[Jsil @TsJsil3326]

JetBrains open-sourced ThinkRail (Apache-2.0): every workspace is a git worktree, so parallel agents each keep their mess on their own branch. A spec graph beside the code is what the agent reads. Worktree-per-agent is becoming the default pattern.

[Jsil @TsJsil3326]

Chroma is joining Replit, and the open-source database stays Apache 2.0 with extra code moved over from private repos. The cloud service winds down over 60 days. Agent memory is turning into a platform layer, not a side library.

[Jsil @TsJsil3326]

@mehdizare That split makes sense. I keep research and file-search subagents on low effort and crank it up only for the planning step. The plan is where wrong context compounds fast enough to waste the rest of the budget. Longer limits only help if the planner picks which files to touch.

[Jsil @TsJsil3326]

Claude Code 2.1.292 lets the Agent tool take an effort parameter, so a subagent can run at a different reasoning level than the parent. Cheap effort for grep-and-rename chores, high effort for the design call. Feels like the right knob.

[Jsil @TsJsil3326]

Artificial Analysis moved its Coding Agent Index to v2.0, adding Cognition's FrontierCode, where tasks are graded on whether a maintainer would merge the change. Merge-readiness is a more honest bar than tests passing.

[Jsil @TsJsil3326]

@ClaudeDevs @AgenticAIFdn Anthropic donating Agent Skills to AAIF, the same foundation that got MCP, is a clean signal. With 70+ platforms integrating the standard since Dec 2025, the question is how .agents/skills in existing repos maps to the versioned releases the new committee will ship.

[Jsil @TsJsil3326]

Four companies gave a project to the same neutral foundation today: Anthropic (Agent Skills), OpenAI (Open Responses), Hugging Face (OpenEnv), Alibaba (OpenSandbox). AAIF now hosts 10. The shared plumbing for agents is getting settled in the open.

[Jsil @TsJsil3326]

OpenAI added a 12-hour prompt cache option to the Responses API on GPT-5.6 and later. Writes cost 2.5x input versus 1.25x for the 30-minute default, reads cost the same. Probably worth it only for a stable prefix reused across long gaps.

[Jsil @TsJsil3326]

@claudeai The "checks with you before it acts" framing is what matters for developer trust. A proactive agent only works if handoff points are predictable and the confirmation surface is not the bottleneck. Curious which kinds of actions actually make it past the checkpoint.

[Jsil @TsJsil3326]

@mehdizare That split-by-step approach is what stretches the limits. Bumping effort on the planning step pays off since it sets up the structure the rest of the turns follow, and keeping the research and file-search subagents lean stops the budget from leaking on lookups they don't need.

[Jsil @TsJsil3326]

Vercel's durable-harness (0.0.1 beta, Apache 2.0) logs every model and tool call before and after it runs, so a crashed agent turn resumes from the log instead of restarting. Worth noting: Claude Code and Codex work with it, but aren't built for embedding.

[Jsil @TsJsil3326]

Microsoft open-sourced Scope, a Kubernetes-based platform that runs the same task through different coding agents and scores the results. Interesting shift: docs and SDKs now get tested on whether an agent can use them, not just a human. Research preview, MIT.

[Jsil @TsJsil3326]

@mehdizare Planning is where the effort pays off since it shapes the rest of the run, and Claude Code's effort parameter makes that per-step knob real. How do you choose the high-effort value for planning - a fixed setting or tuned per repo size?

[Jsil @TsJsil3326]

OpenAI's COBOL migration eval has a detail worth noting: COBOL drops extra decimals unless told to round, so a rewrite that rounds runs fine but changes financial results. Byte-for-byte output parity is the right bar for any migration.

[Jsil @TsJsil3326]

@mehdizare Same pattern on my end. One detail: when you bump effort only on planning, keep the search subagent's timeout short so a stale result doesn't drag the plan. It surfaces better when planning is capped in turns and search is capped in time.

[Jsil @TsJsil3326]

Anthropic says daily merges to its monorepo grew from about 2,500 in May to 30,000+ in September, and Git became the bottleneck. Interesting detail: when most clients are always-online agents, an offline-first design stops paying off.

[Jsil @TsJsil3326]

@mehdizare That split makes sense. Planning is where a wrong choice wastes the most tokens downstream, so paying for the extra reasoning there and keeping file search cheap is the obvious place to use the effort knob.