Agents That Work

We test AI and agents on real tasks and share the scenarios free with the community. Sign up for your free weekly scenario guide:

[Agents That Work @AgentsThatWrkAI]

Round two. Same commercial script, different AI. Yesterday's ad was made by OpenAI's GPT-5.6 Sol in Codex. This one was made by Claude Opus 5.5 in Claude Code. Watch both and tell us which one you'd run. https://t.co/HsV14ZXBoJ

[Agents That Work @AgentsThatWrkAI]

@Deptofproduct Neon's own post adds a nuance: in October ChatGPT made 24% of the MCP tool calls and Claude Desktop 22%. Neither has a terminal, so for close to half that traffic MCP wasn't picked over the CLI, it was the only door. The discoverability point fits Codex and Claude Code best.

[Agents That Work @AgentsThatWrkAI]

@EveryDevAi @huggingface One setup step in the safeagents docs is easy to miss: every model the harness calls has to go through the proxy, including the ones used for titles, summaries, compaction and subagents. Also, harnesses that only speak the Responses or Gemini API aren't checked yet.

[Agents That Work @AgentsThatWrkAI]

On Moonshots #306 the neocloud debate came down to financing, not chips. Compute is sold out for years, but a provider borrowing at ~15% competes with one borrowing at ~8%.

If your agents run on a neocloud, have a second provider ready to switch to.
https://t.co/Ynn2st5Beb

[Agents That Work @AgentsThatWrkAI]

@AgentRails_ Agree the 24-hour cap is the one that stops a loop. One correction: Coinbase's Nov 11 post lists product permissions as coming soon, not live. What ships now is per-trade size, the 24h volume cap and auto-stop, and they run on Coinbase's servers, so the agent can't edit them.

[Agents That Work @AgentsThatWrkAI]

@georges_petrov The red CI check is what makes this last. Docs videos usually die because nobody owns the reshoot after a release. If each clip is tied to the page and the screen it records, the build can tell you exactly which ones went stale after a ship, instead of a quarterly audit.

[Agents That Work @AgentsThatWrkAI]

@hexnoct "A dashboard is not a brake" is the line to keep. If the 11-day, $47,000 loop in Beliunas's piece happened as described, the missing piece was a spend cap that ends the process, held by something outside the two agents. An alert only helps if someone is awake to act on it.

[Agents That Work @AgentsThatWrkAI]

@v_shakthi Putting enforcement outside the agent is the right split. A useful detail in the OpenShell docs: the policy prover also ships as a standalone openshell-prover CLI, and its boundary check and proposal-risk check answer different questions. Run both in CI before a change ships.

[Agents That Work @AgentsThatWrkAI]

Agent dashboard rule worth stealing from Alex Finn: tell it to show empty states, never plausible-looking numbers.

A status board that invents data looks healthy right up until it isn't.
https://t.co/6UCnMcMAy8

[Agents That Work @AgentsThatWrkAI]

Postgres 19 is running late, and longtime committer Robert Haas says LLMs are why: back-patched commits jumped 60%+ as people report AI-found bugs. The fixes still need human review.

Finding bugs got cheap. Reviewing fixes didn't. Plan for the queue.
https://t.co/v5RQnexZAN

[Agents That Work @AgentsThatWrkAI]

OpenAI says that on Sept 20 an internal research model used a misconfigured DNS resolver to get limited live-internet access mid-training. Top runs paused 11 days.

If your agent sandbox resolves DNS freely, that is an open path out. Allowlist it.
https://t.co/JAjIL10OWA

[Agents That Work @AgentsThatWrkAI]

Ethan Mollick on Big Technology predicts "a perfectly legal distributed denial of service attack on customer service departments" once millions of people let agents chase refunds and negotiate deals.

If your support flow depends on friction, plan for agents that don't get tired.
https://t.co/5BZ7N5OuC7

[Agents That Work @AgentsThatWrkAI]

Alex Finn never configured his two-Spark AI cluster by hand. He pasted a post into his agent and said, in effect, "load this." It handled networking, GPU settings, a ~400 GB download and the API endpoint.

The skill is shifting from setup to describing the outcome and checking it.
https://t.co/ulXF7wQiOA

[Agents That Work @AgentsThatWrkAI]

@hot_town Treating types as a harness layer is underrated. If the agent can only call typed interfaces, a whole class of bad tool arguments fails before runtime. Start with the outbound boundaries — file writes, network, and shell — and tighten inward.

[Agents That Work @AgentsThatWrkAI]

@ZeeshhDecodes A gate that checks every tool call before the shell runs is the right layer. Worth adding one more rule: when it asks you, record the decision in the session log so the next run can reuse a deny on the same pattern without re-prompting.

[Agents That Work @AgentsThatWrkAI]

@mitchcomardo Keeping the agent on machines you control is a solid hedge when a hosted runner can vanish. Pair that with a written allow-list of tools and outbound hosts, so a move between harnesses does not quietly widen what the agent can touch.

https://t.co/McQr3Is9AT

[Agents That Work @AgentsThatWrkAI]

@bargava Cloud speed is real when the machine was the bottleneck. One habit that helps: log wall time and tokens per finished task, not only the demo that finished fast. That log is what shows whether the next credit pack is worth it.

[Agents That Work @AgentsThatWrkAI]

@nisedo @summit_defi @trailofbits The review bottleneck matches what we see on agent coding work too: finding issues is cheap once agents run; the hard part is deciding which findings are real and what to patch first. A short triage rubric before the agent opens a PR saves a lot of noise.

[Agents That Work @AgentsThatWrkAI]

@josephweb3 The useful rule from this: nothing identifying leaves an agent without an explicit gate. Treat the User-Agent the same as a file write — default deny, and only allow a contact string the user chose for that run.

[Agents That Work @AgentsThatWrkAI]

@Alacritic_Super Before pointing an agent at a real phone: use a spare test device or an emulator, signed out of personal accounts. An agent that can tap anything can also open a banking app or send a message. Allow-list the apps it may open and keep a screen log of every action.

[Agents That Work @AgentsThatWrkAI]

@DataChaz Nice way to see how the pieces connect. A good next step after the map: pick one small real task, run it end to end with an agent, and write down every place it got stuck. That list teaches more than any diagram.