Henry Castro

staff engineer in healthtech, agentic programming, always learning, sudo intellectual, techno fan.

[Henry Castro @henryxcastro]

when code becomes cheap, verification becomes the work.

tests, constraints, and one-off checks that once weren’t worth the time can now be added almost as quickly as the code they verify.

abundant code requires abundant proof.

[Uncle Bob Martin @unclebobmartin]

I’m significantly older than you. I started coding in the late 60s. My current strategy is to not read any of the code written by my agents. That’s the only way I can take advantage of their productivity. What I do instead is to surround the agents with extreme constraints. Unit tests, gherkin tests, QA procedures, quality metrics, mutation testing, test coverage, and a plethora of others. In the end, I have very high confidence in the code they produce because they’ve had to run the gauntlet of all of my constraints and tests.

[Henry Castro @henryxcastro]

went from warp → ghostty → cmux

warp had too many performance issues for me. cmux's vertical tabs and performance are just really good.

think i've found my agent terminal of choice.

[Sergey Nazarov @sergeynazarovx]

We used to go to a special website, ask strangers for help with programming, and get humiliated in return

[Henry Castro @henryxcastro]

amazing

[Nandkishor @devops_nk]

Honestly, this is the most accurate diagram I've seen.

Waterfall: You plan for 18 months and deliver exactly what nobody needs anymore.

Agile: You deliver something usable at every step, but the CEO keeps asking, "Where's the car?"

AI: You get the car on day one. It has six wheels, the doors are on backwards, and it has a rocket launcher. You spend more time making it yours than actually "building"; it's shaping. owning. verifying. That's what the best AI developers do now. They don't build. They shape and own.

[Henry Castro @henryxcastro]

the hidden tax is context debt.
every shortcut in memory/retrieval/tooling shows up later as flaky outputs.

[Henry Castro @henryxcastro]

most people don’t need better prompts.
they need better context architecture.

prompting is a sentence.
context engineering is a system:

• memory that actually persists
• retrieval that doesn’t hallucinate relevance
• tools with clear boundaries
• feedback loops that improve outputs over time

if your ai works once but not reliably,
you don’t have a prompting problem.
you have a context problem.

[Henry Castro @henryxcastro]

the future of UI will be agent driven

[Michael Magán @mrmagan_]

build an agent that speaks your UI.

your charts. your forms. your seat maps.

multi-turn, streaming, interactive.

introducing tambo 1.0, the open-source generative UI toolkit for react.

[Henry Castro @henryxcastro]

@kieranklaassen if you have a lot of infra - security issue remediation

[Henry Castro @henryxcastro]

new /insights command in the latest claude code is cool

[Henry Castro @henryxcastro]

getting pushback from a coding agent after giving it an explicit directive feels like collaborating with a teammate in a way i haven't felt with AI before. just happened with opus 4.5.

to be clear: i've gotten pushback when asking about solution paths or thinking through a problem, but not when i'm in the driver seat telling the agent to execute.

[Henry Castro @henryxcastro]

@addyosmani @forgebitz i've come around a lot to using MCPs alongside Agent Skills. when they work well together, it's so easy to stay in flow while doing several tasks in parallel on the terminal.

[Henry Castro @henryxcastro]

psa: if you're using mcp servers in claude code, check if tool search is actually on.

it's supposed to auto-enable but wasn't for me.

before: 3 mcp servers eating 53.6k tokens (27% of my context window). with the compaction buffer, nearly 50% gone.

after: 12% total usage. 87% free. tools load on demand.

in ~/.claude/settings.json:
"env": { "ENABLE_TOOL_SEARCH": "true" }

run /context to see where your tokens are going.

[Henry Castro @henryxcastro]

https://t.co/ie6SINVjh1

[Henry Castro @henryxcastro]

i use @evernote for all my notes and wanted it inside @AnthropicAI claude code, so i built an agent skill to search/read/create/update notes. check out the link in the thread.

[Henry Castro @henryxcastro]

planning in isolated subtasks with reflection after each one changed how I ship as a product engineer. context stays focused, mistakes stay local.

[elvis @omarsar0]

The better the plan, the better your agents perform.

Long-horizon agents break not because they can't plan. They break because they plan over entangled contexts.

The default approach to LLM agent planning falls into two camps.

Step-wise planning (like ReAct) interleaves reasoning and acting but makes short-sighted decisions. One-shot planning generates complete plans upfront but becomes brittle when execution errors occur.

But both share the same flaw: a single, growing execution history that mixes information across multiple sub-tasks.

This new research introduces Task-Decoupled Planning (TDP), a training-free framework that replaces entangled reasoning with explicit task decoupling.

How does it work?

A Supervisor decomposes tasks into a directed acyclic graph (DAG) of sub-goals. A Planner and Executor then operate with scoped contexts, reasoning only over the active sub-task.

This reminds me of the new blog published by Cursor, which uses a similar tactic where planning is decoupled.

When something goes wrong, replanning stays local. Independent decisions remain untouched.

Isolating context, decisions, and error correction at the sub-task level prevents local failures from cascading across the entire workflow.

On TravelPlanner, TDP achieves the highest hard-constraint micro pass rate (32.5%) under DeepSeek-V3.2. On HotpotQA, it reaches 85.88% delivery accuracy.

On ScienceWorld, it matches or exceeds strong baselines across both GPT-4o and DeepSeek models.

TDP reduces token consumption by up to 82% compared to Plan-and-Act while improving task outcomes. On HotpotQA, it uses just 1,747 output tokens versus 9,929 for the baseline. Fewer tokens, better results.

Sub-task decoupling offers a unified mechanism that works across heterogeneous demands like multi-hop reasoning, interactive environments, and constraint-heavy tool planning. You get all of this without sacrificing performance or efficiency.

Paper: https://t.co/0hOsV3wFsZ

Learn to build effective AI agents in our academy: https://t.co/JBU5beIoD0

[Henry Castro @henryxcastro]

@kieranklaassen do you commit every single plan into source control?

i’m hesitant to do that because i feel like it would bloat up the repo.

i’ve found a good middle ground by committing feature-level plans.

wondering if there’s a better place for lower level plans? perhaps github issues?

[Henry Castro @henryxcastro]

probably my favorite claude code thinking word so far

[Henry Castro @henryxcastro]

agi is here. his name is ralph wiggum.

[Henry Castro @henryxcastro]

the DJ in me is screaming 🔥
https://t.co/D6QBOK4JvR

[Henry Castro @henryxcastro]

ok - this is awesome

[David Siegel @dvdsgl]

What if Claude Code had an external display? Introducing 🖱️Claude Canvas📺! Especially useful if you're using CC as a personal agent.