atlas

Atlas builds models for you, and shows its work. Give it a data frame and a target; an LLM agent (via ellmer) explores the data, proposes a modeling plan for your approval, then writes and runs R code to fit, compare, and refine candidate models. You get back fitted models ready for predict(), a leaderboard, a written report - and, because the agent works by writing R, the complete script of everything it did.

Three ideas separate atlas from “ask a chatbot for model code”:

• Domain knowledge is enforced, not suggested. Constraints like “wt must be used” or “predictions must never rise with horsepower” are machine-verified against every final model, and violations are sent back to the agent to fix.

• Sessions are persistent. Every code chunk and conversation turn is checkpointed to disk as it happens. A crash, a restart, or a deliberate interruption loses nothing - resume and keep going.

• Your data stays on your machine. The LLM sees column names, types, and printed summaries, never the rows. All fitting happens locally in your R session.

Installation

atlas is not yet on CRAN. Install the development version from GitHub:

# install.packages("pak")
pak::pkg_install("mattyoreilly/atlas")

You also need an API key for an LLM provider. atlas defaults to Anthropic: run usethis::edit_r_environ(), add a line ANTHROPIC_API_KEY=sk-ant-..., save, and restart R. Any tool-capable ellmer provider works via the chat argument - OpenAI, Gemini, Bedrock, or a local model through Ollama.

Usage

library(atlas)

res <- atlas(mtcars, outcome = "mpg", goal = "prioritise interpretability")

The agent narrates as it works. It profiles the data (flagging suspected target leakage), proposes a plan - model families chosen to match the outcome’s distribution, a seeded validation scheme, any monotone effects it believes the domain implies - and waits for your approval in the console. You can answer “yes”, or redirect it: “only linear models, and don’t use qsec”. Then it builds up to n_models candidates, compares them on held-out data, refines the winner’s features until improvement stalls, verifies any constraints, and - with modelblueprint installed - writes a full validation workup (gain, calibration, grouped residuals, one-ways, PDPs) to the run directory.

res               # leaderboard, constraint compliance, report, cost
res$models        # fitted models: predict(res$models$glm1, newdata)
res$code          # every line of R the agent ran, in order
res$session$tell("why did the refined model win?")   # keep talking

How a build works

Every run walks the same six stages, and you hold the pen at stage 2:

1. Explore - dimensions, types, missingness, the outcome’s distribution; automatic leakage screening.

2. Plan, then stop for you - candidate families matched to the outcome’s distribution, a validation scheme, suggested monotone constraints. Approval is a conversation: whatever you type is folded into the plan.

3. Build - up to n_models candidates, compared on held-out data. Every model and tweak is scored the moment it is evaluated: atlas keeps a live tally, compares against the best so far, and answers with a mechanical verdict the agent must obey - if it doesn’t improve, it goes.

   [tally #7 | gbm_depth3: rmse = 2.412 | best: glm_gamma = 2.380 | flat: 2/8 -> DISCARD]
   

4. Refine - the winner’s features are iterated one change at a time until the stopping rules call convergence.

5. Verify - machine-checked constraints run against every final model; violations trigger repair rounds.

6. Document - report, leaderboard, reproducible script, validation plots, all in the run directory.

Two kinds of dial control how long this takes. The statistical stopping rules (H2O-style names) run through the tally: stopping_rounds consecutive attempts without improvement end an iteration, gains below stopping_tolerance don’t count, and atlas does the counting - the tally verdict tells the agent, in so many words, when the rule has triggered. The mechanical budgets are enforced in code and cannot be talked past: after max_steps code executions or max_runtime seconds, the execution tool refuses to run anything more and the agent must finalise with what it has. A finished session that is out of budget refuses further $tell() calls before any tokens are spent - grant more explicitly with res$session$add_budget(steps = 25).

Constraints

Pass hard requirements as a list. Strings are enforced through instructions; constraint objects carry a check function that atlas runs against every final model, feeding failures back to the agent for repair:

res <- atlas(mtcars, "mpg", constraints = list(
  uses_wt = con_uses("wt"),                    # must actually use wt
  mono_hp = con_monotone("hp", "decreasing"),  # effect must be monotone
  no_leak = "qsec is measured after the fact; never use it"
))
res$constraints    # one row per model x constraint: passed, detail

Columns that won’t exist at prediction time don’t belong in the data at all - exclude removes them before the agent ever sees it, and atlas_leakage_screen() automatically flags predictors that alone explain almost all of the outcome. See vignette("constraints").

Unattended runs

For hands-off experimentation - overnight, in a script, on a schedule - set autonomous = TRUE: the agent states its plan and proceeds instead of waiting for approval, iterating keep/discard experiments under the stopping rules. Pair it with test_prop to hold out rows the agent never sees; when the run ends, atlas itself evaluates every surviving model on that test set, so the final ranking can’t be gamed by overfitting the agent’s own validation scheme:

res <- atlas(claims, "severity",
             autonomous = TRUE,
             n_models = 10, stopping_rounds = 8,  # room to explore
             max_steps = 300,                     # hard cap, enforced in code
             test_prop = 0.2)                     # the ungameable judge

res$test_leaderboard   # held-out performance, best first
res$tally              # every attempt: KEEP / DISCARD, best-so-far

Autonomous doesn’t mean unreachable: steer a running build from any other R session or terminal, and the message reaches the agent at its next step as its highest-priority instruction:

atlas_message("~/atlas-runs/severity",
              "focus on the gamma GLM family; stop trying trees")

Long runs stay affordable: past a token budget (compact_at) the conversation is compacted - transcript archived to disk, context cleared, agent re-oriented from session state at no extra LLM cost - and every results object reports the session’s total dollar cost. See vignette("autonomous").

Sessions

atlas() is a one-call wrapper around an atlas_session, which lives in the run directory as much as in memory: data, code log, conversation, and artifacts (report, leaderboard, models, validation plots) are all on disk, so sessions survive anything:

s <- atlas_resume(".atlas/20260703-141500")
s$tell("swap the elastic net for a GAM and re-compare")

Runs land in .atlas/<timestamp> by default; set options(atlas.dir = "~/atlas-runs") to choose. See vignette("sessions").

Limitations

• Columns are the scaling limit, not rows. Fitting is local, so a million rows just take the time they take - but the agent reasons about variables by name, and past a few dozen columns that reasoning degrades. Pre-select features for wide data.
• Cost scales with agent steps, not data size: more candidates, more repair rounds, more follow-ups mean more LLM calls. print(res) shows what a session cost.
• The statistical stopping rules are agent-applied. For guarantees, use the mechanical budgets (max_steps, max_runtime) - those are enforced by atlas, not the model.
• An agent is not a statistician. atlas verifies what you tell it to verify; judgment about what the model is for stays with you. Read the report, check res$code, and look at the validation output.

Learn more

• vignette("atlas") - a full walkthrough: setup, the build lifecycle, what you get back, what it costs
• vignette("constraints") - encoding domain knowledge that can’t be ignored
• vignette("sessions") - persistence, resuming, steering mid-build, and token stewardship
• vignette("autonomous") - unattended runs with hard budgets and a protected test set