AI Field Journal

A Symfony side project used as a practical AI-assisted development sandbox: rapid implementation loops, model experiments, and real infrastructure instead of isolated demos.

A small automation to collect useful signals from Slack and make them easier to revisit.

Joined the program to turn field experiments into internal demos, workshops, and practical feedback on agentic AI.

A dedicated workspace for AI watch, experiments, technical decisions, execution plans, talks, and training material.

A user-level harness for a consistent agent environment across Claude Code, Codex, Gemini, and other compatible tools.

Shared foundation. One source of truth distributes collaboration instructions and reusable skills across agent runtimes instead of maintaining competing copies.

Interoperability. Centralized provider settings, MCP registrations and drift checks; versioned hooks where supported; kept service credentials scoped locally rather than ambient.

Operational memory. Applied the Karpathy LLM Wiki pattern to a Markdown knowledge base readable by agents on demand, with an active work ledger and a context-persistence workflow.

Used .ia/ as a project-local fallback for notes and experiments that must remain untracked, while keeping durable context in explicit documentation. A global Git ignore prevents this private workspace from leaking into a repository by accident.

Set up a private Mac, Windows RTX 5090, and personal-VPS environment for running and comparing local models through Ollama and LM Studio. Tested Qwen, Mistral, Gemma, and DeepSeek families while keeping the Mac as the main agent workstation and the Windows machine as a selectable GPU runtime.

A Symfony benchmark lab for comparing coding agents, models, runtimes, prompts, and acceptance evidence instead of relying on subjective impressions.

An experimental internal Symfony project that turns Reddit field feedback into an auditable dashboard and an LLM-assisted synthesis workflow.

A Symfony and YouTrack assistant used to test tool calls, multi-turn memory, provider switching, authentication, and user-facing agent UX. A LibreChat POC later validated the Windows local-model route through a real chat interface.

Adopted reviewed Markdown execution plans as the default artifact for substantial agent-led work, with milestones and validation rather than chat-only instructions.

Added a reusable handoff skill that writes compact continuation documents in a dedicated location, separating durable resumption context from transient chat history.

Expanded Agent Context with the todo skill (a parseable Markdown index plus a context file per task) and a cross-agent active-work ledger.

Built a reusable competitive-audit skill for source-fresh landscape reviews: several independent research lenses, iterative coverage, and an explicit gap analysis instead of a short list of familiar tools.

Started an experimental internal program around agentic coding on real Symfony/PHP workflows, with quality, tests, review, legacy code, security, and team adoption as the differentiators.

Started maintaining this concise, public-facing companion to project documentation and the knowledge base, so the evolution of the work remains easy to follow.

First captures selected team-channel signals, Google Chat spaces, and meeting notes as local raw evidence, deliberately separate from curated knowledge and ignored by Git.

Then a weekday Codex automation assembles a French morning brief from calendar, personal mail, TODOs, and a local Claude runner for professional mail. The capture routines use local schedules for regular refreshes; their generated artifacts remain private and untracked.

Added a Mac/Windows clipboard bridge alongside the local-model runtime tunnel, making the Mac workstation and Windows GPU machine practical to use as one agent workspace.

A French talk that unpacks the "hundreds of agents in a few hours" myth: what "several agents" really means across parallel sessions, subagents, delegation, agent profiles, and home-grown or off-the-shelf orchestration.

Created and published a French talk on personal and team knowledge as the context layer for AI agents: why context beats ever-larger prompts.

Presented agentic AI to a less technical audience through concrete project workflows, including why MCP is useful only when it solves a real tool need and should not automatically become the centre of an architecture.

Delivered this talk for ESGI students, connecting Symfony development practice with the realities of AI-assisted work.

A local, read-only desktop operations cockpit bringing together project state, local and Windows LLM runtimes, agent activity, scheduled routines, VPS health, knowledge, and current work.

A local-first CLI that dispatches one prompt to several local model runtimes and makes their answers comparable; the first use case was structured backlog estimation.

Evolved SWE-Jobs from one-off runs toward campaigns, adapter health gates, run diagnostics, scoring contracts, and comparable evidence across runtimes.

Designed, but have not yet adopted, a model that separates the UI client, agent host, worker, and model-runtime host. Its intended shape is a Mac client, VPS coordination and workers, and selectable Mac/Windows inference runtimes with explicit state ownership.

Started a repeatable benchmark that separates microphone and capture quality from transcription quality and technical-vocabulary handling.

Captured reusable rules for approachable agent-facing CLIs: useful default commands, grouped help with examples, clear errors, TTY-safe colours, visible progress, and dry-run plans people can trust.

Tested VS Code and Zed as lighter, multi-project complements to PhpStorm. This remains a complement experiment, not a complete PhpStorm replacement.

Introduced a cross-agent Git convention that records the agent, native session, and task in commit trailers, enforced by a hook while Fork-authored human commits stay frictionless and leave no attribution marker in history.

Example trailers:

AI-Agent: Codex
AI-Session: <thread-id or unavailable>
AI-Task: short-task-slug