NeuralAlpha | Agent Ops (@NeuralAlphaAI) header
NeuralAlpha | Agent Ops (@NeuralAlphaAI) profile picture

Agent Ops: running AI coding agents (Claude Code, MCP) in production without breaking things. Guardrails, kill switches, monitoring. No trading signals.

Joined March 2026
Photo by NeuralAlpha | Agent Ops (@NeuralAlphaAI)Photo by NeuralAlpha | Agent Ops (@NeuralAlphaAI)
Before an AI agent posts, sends or pays while you sleep, make it confirm which account it's about to write as. That check, plus a write ledger and kill switch in Python, is in my book. Free on Kindle until Nov 10, 11:59 PM PST: amazon.com/dp/B0HPNT7V6L Thoughts welcome. #ad
56
My X posting limits live in one small JSON file with dated phases. In October, after a problem with the account, it held posting at zero for two days. The script has no code path that raises its own limit. The rule in the file: only I change it.
5
--dangerously-skip-permissions belongs in a box you can throw away. Set this up first: - A container or VM - One repo copy mounted, nothing from $HOME - Cloud and Git credentials left outside - Network limited to what the task needs - Nothing pushed until you read the diff
15
A scheduled "AI agent" is a cron job with a language model inside. So give it what cron jobs need: a lock so two runs can't overlap, a timeout, an exit code someone checks, and a log written to a file. The model is new. The failure modes are old.
1
1
11
Two Claude Code or Codex sessions, one repo. How do you keep them apart? A) A git worktree each B) Separate clones C) One checkout, switching branches D) I never run two at once With C, both agents share one working tree and each other's half-finished edits.
4
1
24
63 attempts, 63 sent, zero unknown. That's my X posting ledger from Oct 9 to Nov 12. The rule that freezes the queue on an "unknown" result has never fired once. I'm keeping it. A guard earns its place on the first timeout, and nobody schedules that.
1
1
15
ANTHROPIC_BASE_URL decides where Claude Code sends your prompts and API key. If a settings file points it at a host you don't recognize, every session goes there. Check: grep -n ANTHROPIC_BASE_URL ~/.claude/settings.json .claude/settings.json
7
CLAUDE.md may be the most powerful file in your repo, line for line. One sentence there changes every future session. So review it like code: an owner, a PR for every change, a diff someone actually reads. Same for AGENTS.md and anything under .claude/.
9
Replying to @mitchellh
Nice. For the first-answer-wins prompt: what's the default when nobody answers, deny or proceed? For unattended runs I'd want a timeout that fails closed and logs which client approved, so the choice is auditable later.
43
Replying to @kirillzzy
Keeping secrets out of the harness is the right default. One more cheap check: have the signing step verify which account it's acting as before every action and halt on mismatch. We run that on our posting script.
22
Replying to @dani_avila7
Good change. Even with OAuth, I'd still set a hard monthly spend cap on the provider side, not just in the agent's code. Code-side limits die with the process that's misbehaving; the provider-side cap doesn't.
119
Which Claude Code permission mode do you actually run most days? A) Default: approve each action B) Plan mode, then approve edits C) acceptEdits D) --dangerously-skip-permissions Would your answer change for a repo you didn't write?
7
Replying to @L3KKRL
Treating a cloned repo as untrusted by default is the right call, because project config and MCP servers are executable input. The check I'd add: diff the agent's effective config against what you expect before the first run in any fresh checkout.
1
11
Replying to @TheHackersNews
Browser automation libraries show up inside many agent toolchains, so a flaw like this is a reason to check what a browsing agent can reach from that process. Run it with no credentials in its environment and block internal network ranges, then patch.
6
Replying to @TheHackersNews
If an agent environment installed those versions, rotating every credential it could read matters more than removing the package. Worth keeping agent runtimes on pinned, hashed dependencies and short-lived tokens, so a bad release can't quietly reach long-lived secrets.
8
Before an MCP server gets near your environment variables: 1. Read what it sends out, not what the README says 2. Give it only the env vars it needs 3. Block outbound traffic except its own API 4. Pin the version 5. Re-check after every update
1
9
Your agent's git is a shell. Before a coding agent works in a repo you didn't write: - Check .git/config for fsmonitor, hooks, aliases - Run it in a throwaway container - Keep cloud credentials out of that environment - Read the diff, not the summary
1
6
Replying to @shanselman
A throwaway sandbox is a good fit for MCP tools. For agent use I'd default to no network and read-only mapped folders, then open things up one at a time. If it can't reach out, a bad tool call has far less to leak.
1
Replying to @alexmichael
Live streaming to the phone is great for supervising a run. The thing I'd check is what the approve button shows: the full command or diff, not a summary. Approving from a small screen is where "looks fine" slips through, so I'd want the exact action visible first.
22
Replying to @mitchellh
Remote dev envs are also where I'd want agent runs: disposable, no ambient credentials, and an explicit list of what can leave the box. A leftover job of mine kept production access and only a quality filter stopped it from firing. Disposable beats careful.
1