Ashish Tyagi

Software engineer in quantitative finance. Interested in AI agents, MCP, evaluation and observability. Sharing notes from hands-on experiments with OSS tools.

[Ashish Tyagi @Ashish_16_2000]

@rabi_guha Great! Definitely going to give it a try. Just curious, how can agents use this to visualize long-term data trends and plan based on those visualizations?

[Ashish Tyagi @Ashish_16_2000]

@marcelotryle Nice. How much latency does the judge add per shell call? I'd run a cheap pattern check first and only ask the model when a command looks destructive.

[Ashish Tyagi @Ashish_16_2000]

@i_mika_el uvx with a pinned version in the install snippet (uvx [email protected]). It skips the system Python fight and users only need uv. Docker as the fallback for people who won't install it.

[Ashish Tyagi @Ashish_16_2000]

@buildwithpb Agree, and the scary part is how clean it looks. What I do now: read the tests before the code. If the tests are weak, the rest of the diff doesn't matter.

[Ashish Tyagi @Ashish_16_2000]

@expo @JI This is the review bottleneck fixed at the right layer: the agent shows the change working before a human looks. I've been building narrated video walkthroughs of PRs for the same reason, watching a change is so much faster than reading it. Curious how it handles flaky UI flows.

[Ashish Tyagi @Ashish_16_2000]

@yarotheslav Makes sense. Keeping one-time, irreversible actions UI-only is a cleaner line.

[Ashish Tyagi @Ashish_16_2000]

@code_wonders @coder_blvck Oh nice, so I can just ask Claude to add a fresh skill to a collection. Trying that this weekend.

[Ashish Tyagi @Ashish_16_2000]

Free Stanford course on building AI agents, and the homework has you write a harness from scratch. Lectures 1 and 2 are already up, so that's my weekend sorted. Saving it for the memory and multi-agent ones too. Know any other courses this hands-on? Send them my way.

[Kai @hqmank]

Stanford is posting its new AI agents course on YouTube for free.

CS 329Z: Engineering AI Agents goes from simple LLM pipelines to autonomous agents. Lectures 1 and 2 are up so far, with more on the way. 13 slide decks are already on the course site.

Lectures I'm waiting for (slides are out):

- Agent design patterns: ReAct, Reflexion and CodeAct, plus context compaction and observation masking, with the Pi harness as a case study
- Agent memory: working, episodic, semantic and procedural memory, skill libraries, and MemGPT-style tiers
- Multi-agent systems: when to use one agent vs many, git worktree isolation, and why multi-agent setups fail

Both homework repos are public. In the first, you build an agent harness from scratch. In the second, you write a benchmark task and run mini-SWE-agent on it.

Work through the whole course and you'll have a solid grasp of how to build an agent and test it.

[Ashish Tyagi @Ashish_16_2000]

@StackDhruv For internal servers I'd want this as traces per tool call: who called it, what failed. Tools nobody calls are worth cutting too, agents pick better from a shorter list.

[Ashish Tyagi @Ashish_16_2000]

@Aynur_Atabeyli Codebase exploration, easily. It burns the most context and the main session only needs the summary.

[Ashish Tyagi @Ashish_16_2000]

@Corporate_CJP Not dumb at all. Writing code got cheap, owning it didn't. Someone still has to know why it exists and what breaks if it's wrong, and the model won't be the one paged at 3am.

[Ashish Tyagi @Ashish_16_2000]

@jakubjurovych Agree, but I think the inventory moves rather than disappears. Nobody will care how many subagents ran, only which credentials and repos they touched.

[Ashish Tyagi @Ashish_16_2000]

@Josephyala "What remains unverified" is the one I'd make mandatory. Agents are great at listing what they did and quiet about what they skipped. A PR field the agent has to fill, plus tests it can't edit, removes most of the reconstruction work.

[Ashish Tyagi @Ashish_16_2000]

@johncrickett Love the idea of the IDE as a shared space instead of a chat handoff. I'm going to try letting the agent drive the debugger over DAP so we both attach to the same session. Not sure it holds up on real bugs yet, but it beats describing variable values in chat.

[Ashish Tyagi @Ashish_16_2000]

@orcdev @radiumcoders @heyandras Nice that MCP came in as a community PR. Curious what it exposes first, read-only tools or can agents actually edit?

[Ashish Tyagi @Ashish_16_2000]

@FlooWerner @markokraemer @enroute_sh Thanks for sharing! 7,600+ concurrent pushes per second is seriously impressive. Definitely excited to give this a try!

[Ashish Tyagi @Ashish_16_2000]

@shcallaway @cvedotnow Free with no API key makes this easy to drop into CI. Going to try check_stack on uv.lock as a merge gate. Does the MCP path keep the package list, like uploads do?