Fouad

I trace AI agents for a living. Claude Code daily. Views my own.
AI Observability @datadoghq

[Fouad @wfouaaad]

@kentcdodds @steveruizok the weird transitions nobody designed are usually the ones someone depends on, and to an agent they look exactly like bugs

[Fouad @wfouaaad]

game developer has the rule text and anthropic's confirmation: https://t.co/13T23MyB8y

the july issue with the heron_brook resolver and the AgentTool line: https://t.co/zwZuE9NHBh

[Fouad @wfouaaad]

claude code's no-game-clones rule landed oct 27 in heron_brook, a prompt slot filled from server config. no off switch

so pinning your cli won't freeze the prompt. in july that slot told opus 5 not to call the AgentTool unless asked. i'd want that in a changelog #ClaudeCode

[Fouad @wfouaaad]

@swyx llm side has a catch tho, the written reasoning isn't always what drove the answer. cot faithfulness papers keep finding models rationalizing after the fact

[Fouad @wfouaaad]

@simonw if the repo squash merges with title + description as the commit message, the raw <details> tags and the whole wall end up in git log anyway

[Fouad @wfouaaad]

docs: https://t.co/TQcbCjpO3G
release note, nov 2: https://t.co/GXK3kmpr9M

[Fouad @wfouaaad]

anthropic's message threads are pitched at long-running agents, but a thread dies 5 min after its last request

so if one tool call runs long, like a slow test suite, the continue 404s and you resend everything. i'd treat the replay path as the main path tbh #ClaudeAPI #AIAgents

[Dylan Garcia @_dylanga]

it's the PERFECT time

probably 80% of my MCP usage with Inspect is using Datadog. everybody who uses Datadog should use it.

[Fouad @wfouaaad]

security boulevard's write-up (nov 3), with what the engineer and railway's ceo said: https://t.co/FzLdcc0sOp

[Fouad @wfouaaad]

a railway engineer's testing agents went after a rival host. per primcloud's founder they hit the cloud metadata endpoint, asked for every customer's secrets and port-scanned the network twice

the prompt said testing. scope has to live outside the prompt #AIagents #infosec

[Fouad @wfouaaad]

@dexhorthy agree on blocks. most useful one is boring tbh, a verify script the repo owns that exits 0 or doesn't

[Fouad @wfouaaad]

@unclebobmartin ask the agent why its code misbehaves and you get a very confident story. sometimes it's even the right one

[Fouad @wfouaaad]

the auto-merge part is in their execution doc: https://t.co/cy6ydVeAbt
repo, posted oct 31 as a research prototype: https://t.co/AUFz9EMawt

[Fouad @wfouaaad]

deepseek's superagent prototype can auto-merge its PRs under 200 lines once approvals cover the head commit. per the docs that's just prompt text, the host doesn't decide merges

i wouldn't trust a cap that only lives in a prompt. glad they say to use branch protection #DeepSeek

[Fouad @wfouaaad]

@housecor agreed, as long as level 0 lives outside the model. branch protection and no push creds for the agent hold up. a "never push to main" line in the prompt is exactly what the injected text argues with

[Fouad @wfouaaad]

with Datadog Agent Observability, we don't stop at showing you your traces. we improve your agent based on our learnings from using our product to improve our own agent behind it

[Rashel Hoover @RashelHoover]

We pointed Insights at the agent behind Insights.

It found oversized tool results, prompt cache misses and failing tool calls, all backed by trace evidence. Then Bits wrote the PR.

Same loop is available for your agents.
https://t.co/EBHetXdBZA

[Fouad @wfouaaad]

docstrings are in openai-python 3.42.0 from oct 31: https://t.co/uYDpMlbJUE
openai devs thread on what shipped this week: https://t.co/a092bXBnfD

[OpenAI Developers @OpenAIDevs]

Building with the Agents API?

Here's what we shipped this week for you:

[Fouad @wfouaaad]

openai's agents api has auto_review now, but only for new root turns. enable it mid-session and the running turn plus its subagents keep their old mode

that'll bite people, the long turn doing real work stays unreviewed. "Deny manual approval fallbacks" needs docs too #OpenAI

[Fouad @wfouaaad]

robert haas, oct 30: https://t.co/L17tX4Cptc
hn thread: https://t.co/WmyLtOsKMb

[Fouad @wfouaaad]

robert haas on why postgres 19 is late: "LLMs happened."

bugs got way easier to find than to fix, so back-patched commits jumped 60%+ in a year. fine by me tbh, i'd rather wait till november than get rubber-stamped llm patches in back branches #PostgreSQL