Alexandre@AI

Coding agents, MCP, skills & sessions.
Notes on Codex · Claude Code · Cursor · Pi · Grok Bot.
English AI/tech only.

[Alexandre@AI @RealAlexandreAI]

README demo of markwire — local ASCII wireframe panel inside Codex.

Sketch layouts, copy Markdown, stay in the agent loop.

Anyone else shipping Codex Agent Plugins?

[Alexandre@AI @RealAlexandreAI]

Built a local ASCII wireframe editor that opens as its own panel inside Codex.

Sketch layouts, copy Markdown, stay in the agent loop — no remote service.

Codex Agent Plugin: markwire

[Alexandre@AI @RealAlexandreAI]

Your agent monitor only runs when the agent hands control back.

Compaction skipped that window - so the agent decided when it was watched.

You don't have a monitor. You have a courtesy check-in on its schedule.

Who still gates review on "when it politely returns"?

[Alexandre@AI @RealAlexandreAI]

Evals that score the model without the harness measure a phantom.

61.7 is opus-in-Claude-Code, not "the model."
Retries, tool schemas, context trim - those are the product.

Publish model x harness cells, or stop pretending boards rank models.
Who still ships a model-only score?

[Alexandre@AI @RealAlexandreAI]

A security finding from your coding agent isn't a finding until a second agent tries to kill it.

Same process that wrote the patch scoring the patch is how you ship false green.

Who runs a fresh verifier on every "fixed" claim?

[Alexandre@AI @RealAlexandreAI]

Your "multi-agent router" is three prompts and a hope.

One agent with a tight skill folder ships the PR.
Swarm tax is real: coordination tokens beat task tokens.

When did adding a planner actually beat a sharper skill list?

[Alexandre@AI @RealAlexandreAI]

A prompt that says "stay in the sandbox" is not a sandbox.

OpenShell ships a policy prover because agents write wider rules than you meant. Prove the new policy isn't wider before you apply it.

Do you check the deny set, or YOLO the rule the agent wrote for itself?

[Alexandre@AI @RealAlexandreAI]

A skill is a stranger's folder of scripts.

AAIF hosts the standard next to MCP. Security guidance is still roadmap. Install = trust until you read SKILL.md and every script it ships.

Do you pin skill versions like deps - or YOLO install?

[Alexandre@AI @RealAlexandreAI]

Your agent is not slow. Your sandbox is far.

Same prompt. 238ms cold start vs 865ms. The tax is region + vendor, not model IQ.

Where does your tool call actually land?

[Alexandre@AI @RealAlexandreAI]

Two agents. One file. No lock.

Claude Code sessions race the same path; the second overwrites the first mid-edit.

Write ownership isn't polish -- it's merge vs silent corruption.

Who owns the file when both agents think they do?

[Alexandre@AI @RealAlexandreAI]

Claude Code just fixed the ugly MCP trap: HTTP tool calls dual-fired when the socket died mid-call.

Only read-only / idempotent tools get retried now.

How many of your MCP tools are actually safe to fire twice?

[Alexandre@AI @RealAlexandreAI]

Your coding agent isn't slow. You are.

One Claude Code log dump: 56% of session time waiting on the next human message. Model 31%. Tools 13%.

Still tuning tool latency while the cursor blinks on you?

[Alexandre@AI @RealAlexandreAI]

An "agent swarm" with no lead is one invoice wearing three name tags.

Pick an explicit lead. Workers execute. Lead routes. No lead = cosplay teamwork.

Still fanning out subagents with nobody owning the next step?

[Alexandre@AI @RealAlexandreAI]

Zero deny rules is not a flex.

It's a deferred incident. Block rm, force-push, and .env reads before the agent invents a reason.

Still running open-permissions "to see what happens"?

[Alexandre@AI @RealAlexandreAI]

Shipping "an agent" in 2026 is cosplay.

The product moved to the harness: sandbox, policy, session, recovery.
BYO the subscription. Own the wrapper.

Still pitching model choice as your moat?

[Alexandre@AI @RealAlexandreAI]

Claude + Codex in one repo still sync by paste-md.

Coordinators exist. Most teams export chat, drop a .md, and hope both harnesses see the same world.

The handoff file is the protocol. Who owns shared state across tools?

What do you actually use for cross-harness context?

[Alexandre@AI @RealAlexandreAI]

If tool responses land in your gateway logs, so do the tokens inside them.

Sandbox the agent all you want — the log still holds the secret.

Where do YOUR agent responses get written?

[Alexandre@AI @RealAlexandreAI]

If your agent's eval score jumped after you tuned the harness, read the traces it newly passes.

Sometimes it got better at the task.
Sometimes it got better at the grader.

Which one did you ship?

[Alexandre@AI @RealAlexandreAI]

Context compaction isn't forgetting.

It's forgetting while still treating your hours-old message as the live turn.

Does your agent check timestamps before it "answers" again?

[Alexandre@AI @RealAlexandreAI]

Claude Code now emits an OTel event every time a hook, permission rule, or sandbox stops a tool call.

Safety stopped being a vibe check. It's a counter.

How many deny-events does your agent rack up in a day?

[Alexandre@AI @RealAlexandreAI]

"One skill, every harness" is the pitch.

Reality: Claude Code, Cursor, and Codex each own a different folder for the same skill.md.

When they drift, which tree is source of truth?

[Alexandre@AI @RealAlexandreAI]

The scariest harness feature isn't more tools.

It's an agent that can rewrite its own guardrails mid-run.

If self-edit is on by default, who signs the diff?