Tag: AI Agent
All the articles with the tag "AI Agent".
-
Fiber's Double Buffering Teaches Agent State Three Moves: workInProgress, Commit, Rollback
blog205 gave the 3 principles Fiber teaches agent state design; this post takes 'double buffering' apart to the bone: workInProgress dry runs, atomic commit switches, one-command rollback discards — mapping to the agent's draft isolation, atomic gate flips, and zero-cost rollback. The answer has been sitting in Fiber for 9 years. This is fe2agent extra C, closing out all 13 posts of the series.
-
The Agent's UI Is Frontend's Real Home Field: Chat Is Just the Fallback
From Cursor to ChatGPT, agent products have all grown into the same list of message bubbles — not frontend's destiny, just design laziness. Chat is the lowest-cost UI for building trust with uncertain output; once trust is established, it should give way to tool panels, state observation layers, and decision gates. Extra B of the fe2agent series (From useEffect to Agent Loop).
-
MCP: The npm of Agent-Land — and Frontend May Have to Pay the Supply Chain's Old Bill Again
Ten years ago, 11 lines of code got unpublished and half the JS ecosystem went down; ten years later, MCP has grown into the npm of agent-land, and frontend may have to pay the same bill again. This is fe2agent extra A: from left-pad to typosquatting, to the first faceplant of my chrome-devtools-in-cron setup, and to postmark-mcp, the first MCP supply-chain case already on the books.
-
From Sweating the Bundle to Sweating the Tokens: Shipping the Agent to Production, Then Giving It a Face
Frontend ships by sweating bundle size; agents ship by sweating the token bill — same craft either way: put a visible gate on invisible growth. Part 10 (the finale) of the fe2agent series — tracing, three-layer budget breakers, checkpoints, and distribution-drift checkups, with tiny-agent v1.0 (trace.js three-column cost ledger + dollar circuit breaker + checkpoint/--resume), plus a replay of that $4200 night through three gates.
-
Flaky Isn't a Bug Anymore, It's the Physics: Migrating Your Frontend Testing Mind to Evals
In frontend a flaky test is a bug to fix; in agent-land flaky is the physics — run the same task 10 times and get 7 perfect, 2 verbose, 1 confidently wrong. Exact match is dead on arrival, a single run proves nothing, and your eval set is your scar collection. Part 9 of the fe2agent series, with tiny-agent v0.9 (eval case set + multi-run pass rates + LLM judge).
-
A Subagent Is Just a Web Worker: If the Brief Was Vague, Don't Blame the Contractor
You think spawning a subagent is starting a thread; really you're hiring a temp — it sees none of the parent agent's history, and every assumption missing from the brief gets filled in by guesswork. A task description is a lossy, schema-free serialization. Part 8 of the fe2agent series, with tiny-agent v0.8 (spawnSubagent + a delegate_task tool + child-loop accounting).
-
Pop a window.confirm at the AI: Permissions, Hooks, and Human-in-the-Loop
blog06 covered "how to recover when it goes wrong"; this one covers "even when it's right, it doesn't get to just do it" — an rm -rf that succeeds on all three retries is the real disaster. Permission is a spectrum, not a switch; confirmations must name consequences; and a denial is information you owe back to the model. Part 7 of the fe2agent series, with tiny-agent v0.7 (danger-level table + confirmation gate + JSONL audit log).
-
In Agent Land, Errors Are Scheduled Work: From Error Boundary to Verifier
The model confidently declares "done" while the work isn't — an error that doesn't error, and no layer of try/catch will catch it. Mechanical failures get retry with exponential backoff plus a hard max_steps brake; semantic failures get an independent verifier — and verification must run in an isolated context. Part 6 of the fe2agent series, with tiny-agent v0.6 (withRetry + MAX_STEPS + verifier).