updown move enter open esc close matheesan.ca / index
    // not just voice agents

    The rest of the job is still software engineering: API reviews, state machines, integrations, infra, tests, release trains and incidents. The tools change; the review bar doesn’t.

    click a stage to filter the timeline · click again to clear

    Tooling timeline

    2021 autocomplete The editor starts suggesting.

    GitHub Copilot

    build

    Inline suggestions, right in the editor.

    Boilerplate, tests and unfamiliar APIs — every line reviewed.

    OpenAI Codex

    build

    Natural language to code before chat felt normal.

    Small prototypes and interface sketches; the spec became the prompt.

    Hugging Face

    research

    The model shelf: transformers, tokenizers, datasets.

    First stop for NLP prototypes and model comparisons.

    2022 the generative breakout Chat, speech and the first chains.

    ChatGPT

    research

    A general model behind a dialogue box.

    Debugging partner and rubber duck; replaced tab-hoarding, never replaced a repro.

    Whisper

    build

    Open speech-to-text that ran locally.

    Transcribed calls and audio notes; made voice-agent debugging searchable.

    LangChain

    build

    Chains, tools and the first RAG patterns.

    Prototyped retrieval and tool flows; learned where the abstraction earns its keep.

    2023 editor-native + local AI moves into the editor and onto my GPU.

    Cursor

    build

    An AI-native editor with multi-file context.

    Moved from autocomplete to reviewing diffs across a repo.

    Ollama

    operate

    Local models on hardware I control.

    Private experiments on the RTX A2000; no data left the lab.

    AutoGPT

    build

    The first runaway agent loop.

    Learned to cap iterations, scope tools narrowly, and verify before acting.

    2024 context + protocol Long context, standard tools, routing + retrieval.

    Claude 3.5 Sonnet

    review

    Long-context code reasoning that held a module.

    Refactors, reviews and test plans; output still went through a diff.

    MCP

    integrate

    A standard way to give models tools and data.

    Wrapped APIs and docs as tool servers; least privilege, read-only by default.

    LLM routers + RAG

    integrate

    Provider fallback plus retrieval, not just a bigger prompt.

    Routed across providers, grounded answers in a knowledge base, and measured retrieval separately from generation.

    2025 agentic coding The terminal becomes the agent surface.

    Claude Code

    build

    A terminal agent with repo context.

    Skills, subagents and repeatable workflows; the agent works where the code lives.

    OpenAI Codex

    build

    Parallel code tasks in the cloud and CLI.

    Independent patches and test scaffolding, reviewed like a junior PR.

    Gemini CLI

    research

    Large-context repo analysis in the terminal.

    Migration sweeps and “where is this used?” questions across big repos.

    2026 the protocol layer The tool protocol settles; workflows get packaged.

    MCP servers

    integrate

    Tools, data and internal APIs behind one interface.

    Standard tool surface for agents; scoped credentials and audit-friendly boundaries.

    Claude Code skills

    build

    Packaged, reusable agent workflows.

    API contract generation, test writing, release notes; guardrails in code.

    Ollama + Open WebUI

    operate

    Private inference with a usable front end.

    Local models for sensitive experiments and offline eval runs.

    optimize for

    short feedback loops · visible evidence · least privilege

    do not trust

    green tests without counts · agent output without a diff review · benchmarks without my workload

    daily stack

    Claude Code + MCP · Cursor/Copilot inline · Ollama for private · an eval harness for proof

    18 tools · 6 years · one loop selected from the tools that changed the loop, not a market ranking