Ask HN: What are you working on? (August 2026)

Working on an Agentic IDE (free) - multiple AI agents support + classic IDE experience

https://shikigami.dev

BETA - means, production ready, but not still extending and experimenting with features - I keep working on it, and collecting feedback.

I have a Pebble Time 2, which is an open source watch. I built a voice operated AI assistant for it.

It's hosted on a VPS, connected over Tailscale, and has a realtime TTS server (Piper).

The Pebble has built in transcription, so my app takes the transcript, hits the server endpoint with three possible paths:

1. Deterministic response

The server looks for certain triggers from the transcript like "Add task," "Complete task," "Set a timer," etc that will run a simple function on server and write / retrieve to a database, no LLM call needed. It then responds with the TTS output and the watch speaks. This path is nearly instant.

2. Fast response

If no trigger is found, then it hits a gpt5.6 Luna endpoint, which decides if it can answer the query itself, or if it needs to defer to the big brother model. If it can handle a query itself, it usually responds to simple questions in less than 8 seconds. It's also good about properly deferring, very close to 100% after a few weeks.

3. Deferred response

The VPS hosts a Hermes installation, and for more complicated tasks or for ongoing projects the Luna router sends it to the Hermes agent. This is typically quite slow, as you'd expect, but remarkably capable. I built the voice app to cache a few lines like "this will take some time" which will automatically trigger if the router indicates it's deferring, and then will respond with the full response from Hermes once it's finished (sometimes minutes later).

Pebble has a quiet mode, and my app will only display the text response if it's on, otherwise all responses are run through the TTS. Piper is so fast that this adds almost no time to the response.

In only a few weeks this has become the primary way I set up tasks, review and close them, set timers and reminders, log my weight, check in on the progress of projects I'm running on Hermes, and ask simple day to day AI questions.

Since Pebble is open source, AI tools are very good at building it, and it's trivially easy to set it up so that you can just ask the model to deploy it live. It's an awesome experience to tell it to change something, and two minutes later, that change is on your wrist.

Very fun project!

Been building a data ontology project i call Tirith - think palantir's gotham + their different modules on top. Can ingest from different sources (textual, structured or freeform for now, images + video coming maybe later), detect relationships between objects, view graph, geospatial view and rudimentary analysts view. Very raw but with the help of nlp and/or openai compatible cheap model like deepseek its already amazingly capable.

And second thing im building is out of sheer frustration as a father of 5 in the school info management - https://www.classtable.org - aims to be a full school management software which is easy to use for both school staff, students AND (personal itch) for parents to keep an eye on their childrens progress

I'm still working on Hackerman recorder. As a programmer, I really enjoy having live coding videos playing in the background while I work. But it's surprisingly difficult to find hour+ long videos of coding without commentary on Youtube.

What I really wanted was an infinite coding video, so I built one. Pick your IDE/monitor/ambience of choice, hit the Screensaver button, put it on a second monitor or TV, and enjoy infinite coding ASMR.

Recently added a Paper Mode which shows an animated quill that scrawls across a parchment for public domain novels.

https://hackerman.specr.net

I’m currently on the beach at Google (since mid July), so I’m looking internally and externally for a new role. But I’m exhausted, so I’m playing around with local LLMs and DeepSeek more as well. I’m currently trying to get an agent harness going where orchestration happens via bazel, and agents are completely isolated beyond what context you give them access to as expressed in starlark. Agent executions are organized as a DAG like a build would be, where the output of one agent can serve as context for another (there is also a feedback mechanism which is needed for arbiter nodes).

I’ve mentioned it before, but we can get some pretty good results for developing implementation and tests with this setup. The context isolation makes it easy to separate test and code writing using a specification as a source of truth. Tests are then run by an arbiter that feedback to tests or implementation depending on who it blames for test failures. The tests are then treated as an independent implementation, and through the math of coincident errors, the arbiter feedback loop ensures that both are eventually correct (N-versions and clean room research did this with humans in the 80s). This then allows us to use less reliable LLMs to develop code, like Qwen on a MacBook Pro.

I hope to have something out on GitHub before my beach period ends (well, if I’m unsuccessful in getting a new role, then I’ll have a lot more time to play at least).