Skip to content

Roadmap

Everything below shipped on 2026-08-29, the rename included: the package and catalogue id are now ai_agent, with the display name “AI Agent”.

What comes next, in order. The list was last revised after comparing the plugin against the QGIS plugin catalogue — in particular the GeoGPT AI Agent listing, which occupies the old qgis_ai_agent slug. That plugin has no released version and no visible source, so the comparison is against its claims, not its code; the claims still mark out territory users expect an agent to cover.

First-release blocker resolved

  1. Done — renamed to ai_agent. The distinct package and catalogue id avoid the existing qgis_ai_agent listing. The first public submission is 0.1.0, non-experimental.

Shipped 2026-08-29

  1. Done — analysis recipes in the skills. Clustering, hotspot detection and regression are already reachable through run_processing, but the model has to guess the right algorithm chain. A recipes section in skills/processing/SKILL.md — named task, algorithm chain, caveats — turns "find hotspots of cafes" into one pass instead of an exploration.
  2. Done — terrain and interpolation guidance. Same shape as above: DEM slope and hillshade, TIN and IDW interpolation as documented chains over the existing processing tools.
  3. Done (narrow) — a 3D view domain. open_3d_view opens a named view of the current layers when the QGIS build exposes that API. Terrain, camera and exaggeration still use the visible, separately approved run_python escape hatch.
  4. Done — annotations domain. Text notes and markers live in the project's main annotation layer; they can be listed and removed by id.
  5. Done — a run journal. Applied runs write a plaintext Markdown summary to ai_agent_runs in the active QGIS profile and report its path. Owner-only permissions protect the directory and files where supported; the privacy page documents the contents, persistence and cleanup limits.
  6. Done — a Gemini preset through its OpenAI-compatible endpoint. It is available alongside the other provider presets and remains on the common OpenAI dialect.

Added along the way

  • A web domain — search_web (with a loud Wikipedia fallback on locked-down networks), fetch_url and geocode; the geocoded bbox feeds download_osm directly.

Agent quality — audit 2026-09-03

Found by reading the core end to end, ordered by expected effect on how well the agent works:

  1. Done 2026-09-03 — retries in the transport. A single 429, 502/503 or a dropped connection ends the run with an error; core/llm/client.py and stream_runner.py make exactly one attempt. Bounded retry with backoff on 429/5xx and connection resets, with a "retrying…" line in the feed.
  2. Done 2026-09-03 — prompt caching on Anthropic. The system prompt and the tool schemas are identical between the turns of one run and are billed again every turn. Mark the static prefix with cache_control in llm/anthropic.py; keep the dynamic tail (plan, queued steps, project context) last so prefix caching on OpenAI-style endpoints hits too.
  3. Done 2026-09-03 — the starting project context was thin — layer names and geometry only, so nearly every run spends its first turn on list_layers. Add CRS, the selected-feature count and, where cheap, feature counts; stay within metadata-class data.
  4. Skill loading cost a turn per domain. Shipped now: the prompt asks for one-turn loading and /skill preloads before the first turn. Next: let load_skill take a list of names.
  5. No evaluation harness. Prompt and skill edits are judged by feel. Record real transcripts (tool calls with their results) and replay them against the loop with a scripted model — a regression suite for the agent's behaviour, and the biggest single lever for "it must work well".
  6. Verification always runs the full round. A trivial single-step non-visual change still costs a 3–4-turn verification; skip or shorten it when the batch had one step that a single read can confirm.
  7. Compaction drops, never summarises. After KEEP_FULL_RESULTS older results shrink to one-line notes; a summary of the dropped span would keep forty-turn runs coherent.
  8. The journal has no numbers. It records steps, not tokens or wall time per turn — the two figures that show where a slow run went.

Considered and set aside

  • QField / mobile integration — a different product with its own runtime; nothing in the agent loop transfers.
  • Multi-user collaboration — the agent is a single-user tool by design; collaboration belongs to the project storage, not to the assistant.
  • Country-specific statistics integrations (population forecasting, national statistics APIs) — data sources, not agent capabilities; the general answer is add_service_layer, run_python and, where a service speaks HTTP, a user-side script.