AI Agents:
Done Right.

• Single-prompt, single-response
• No memory across sessions
• No tools or external data calls
• One-shot text generation
• You fix errors manually

• Plans & executes multi-step tasks
• Uses tools: code, web, files, APIs
• Iterates, self-corrects on errors
• Can spawn sub-agents (multi-agent)
• Example: write → run → debug → plot
  entire SED fitting pipeline

Fixed roles + Standardised Operating Procedures. Assembly-line workflow (manager → architect → engineer) produces structured artefacts, reducing cascading hallucinations from naive LLM chaining.

"Conversation programming": customisable agents interact through structured dialogues, each combining LLMs, human input, or tools. Spans maths, coding, operations research.

Planner↔reviewer loop → control agent executes. Four building blocks: planner, memory module, action space, verifier. Removes human-in-the-loop from routine steps while preserving structured oversight.

Prompt sampler → LLM generates code diffs → evaluators score → database updated iteratively. No human-authored templates. AI Scientist-v2: first fully AI-generated peer-review-accepted workshop paper.

Tab-complete as you type. Context-aware completions for astropy, numpy, matplotlib.

Ask questions, explain code, generate functions, fix bugs. In sidebar or inline.

Multi-file, multi-step tasks. Runs terminal, self-heals errors. GA on VS Code & JetBrains.

Assign a GitHub Issue → Copilot opens a PR autonomously. You review and merge.

from dustpy import Simulation
from dustpy import constants as c

sim = Simulation()
sim.ini.star.M = 1.0 * c.M_sun
sim.ini.disk.alpha = 1.e-3
# ↑ Copilot completes
#   the ini block from
#   context.

1. You ask ChatGPT: "find papers on gap opening criterion"
2. ChatGPT invents 3 plausible-looking bibcodes (hallucination)
3. You search ADS manually for each one — 2 don't exist
4. You copy the real BibTeX by hand into your .bib file
5. You paste the BibTeX back to ChatGPT for the next question
6. Total: ~10 min per reference · error-prone · not reproducible

1. You type: "find papers on gap opening criterion"
2. Copilot calls ads_search() via the ADS MCP server
3. ADS returns real bibcodes, titles, abstracts
4. Copilot calls ads_export() → verified BibTeX
5. Agent appends to paper/bibliography.bib automatically
6. Total: ~5 sec · zero hallucinations · logged in prompts/

• search_papers — full Solr syntax query
• get_record — full metadata by bibcode
• get_bibtex — export BibTeX directly
• get_citations — who cited this paper?
• get_references — what does it cite?
• citation_metrics — h-index, i10, totals
• manage_library — reading lists

• Search papers by keyword / author / year
• Get abstracts, arXiv IDs, DOIs
• Export BibTeX for individual papers
• Citation metrics (h-index etc.)
• Manage personal ADS libraries
• Natural language paper discovery
• Works with Claude Desktop & VS Code

Even temperature=0 doesn't guarantee identical outputs across model versions. Copilot's base model changes silently.

Copilot reads your whole repo. The same prompt with a different codebase gives different code. Context is invisible.

GPT-4o of today ≠ GPT-4o of six months ago. Updates happen without notification. 'I used Copilot' isn't enough.

If you don't record your prompts, no one — including future-you — can reproduce your AI-assisted analysis.

Authors remain solely responsible for all content, regardless of how AI was used. AI tools cannot satisfy authorship criteria.

Disclose the tool, version, purpose, and section. The threshold is content generation, not volume of use.

LLMs cannot approve manuscripts, hold copyright, or be accountable. They may not appear in any author list.

• AI cannot be listed as author
• Disclose in Methods OR Acknowledgements
• Cover letter disclosure also expected
• No explicit copy-editing exemption (err on side of disclosure)
• No explicit AI image ban (but integrity rules apply)
• Oxford Academic screens for NLP tools in submitted author lists

• AI cannot be listed as author
• Disclose in Methods section (preferred)
• Copy-editing AI EXEMPT — explicit policy (grammar, clarity, language only)
• AI-generated code = AI content: must be disclosed
• Authors verify all AI-assisted content for accuracy

• AI cannot be listed as author
• Disclose in Methods (required) — not in Acknowledgements alone
• AI-generated IMAGES are BANNED: figures, graphical abstracts, all
• Copy-editing AI exempt (Grammarly-style)
• Specific tool + version + purpose required
• Reviewers may NOT upload manuscripts to any AI tool

• GitHub Copilot — free for students & verified faculty via GitHub Education (unlimited chat + agent mode)
• Claude.ai — free tier (Claude Sonnet, limited messages / day)
• ChatGPT — free tier (GPT-4o, rate-limited)
• Gemini — free via Google AI Studio (1M context)
• Ollama — free, fully local, runs on your laptop

• Copilot Pro €10 / mo — higher premium-request quota
• Claude Max €18 / mo — unlimited Sonnet, Projects, Opus access
• OpenAI Plus €20 / mo — GPT-5, o4-mini, DALL·E
• Anthropic API ~€3 / M tokens (Sonnet), ~€15 (Opus)
• OpenAI API ~€2.50 / M tokens (GPT-4o)

• Frame as reproducible research infrastructure — like compute time on LRZ
• Typical researcher: €20–50 / month covers heavy daily use
• Ask IT: check for LRZ / TUM / LMU institutional licence (Microsoft M365 Copilot)
• Grant line: "AI-assisted analysis tools" in DFG / ERC proposals is accepted
• Template: "I request €X/month for AI API access to support reproducible research in [domain]…"

• 5 domains: disk · cosmo · retrieval · LSS/MW · X-ray
• copilot-instructions.md — your lab's conventions, once
• AGENTS.md — project-specific context per repo
• ARCHITECTURE.md — agent roster & 9-stage pipeline diagram
• pipeline-agent.agent.md — 9-stage research pipeline orchestrator
• hypothesis-agent.agent.md — ranked hypothesis generation + ADS
• analytical-agent.agent.md — SymPy pre-analysis & scale estimates
• setup-agent.agent.md — config files + SLURM/PBS job scripts
• analysis-agent.agent.md — simulation post-processing & figures
• interpretation-agent.agent.md — physical interpretation, Gate 2
• spectral-agent.agent.md — X-ray spectral fitting (Sherpa)
• mcmc-agent.agent.md — emcee posteriors, 68% CIs
• paper-agent.agent.md — LaTeX manuscript + ADS citations + auto-review
• literature-agent.agent.md — ADS-verified citations
• simulation-agent.agent.md — FARGO3D/GADGET sims
• retrieval-agent.agent.md — petitRADTRANS retrievals
• Skills: dustpy · fargo3d · pluto · radmc3d · yt · sherpa
• envs/base.yml — conda env, pinned versions
• .pre-commit-config.yaml — style + BibTeX doi= checks
• CITATION.cff — GitHub 'Cite this repository'

• Extensions pre-configured via .vscode/extensions.json — install once, share with team
• ADS MCP pre-wired — Copilot fetches verified BibTeX, no hallucinated DOIs
• pre-commit + CI enforce style, file-size limits, and doi= on every commit
• CITATION.cff triggers GitHub's 'Cite this repository' button
• Zenodo and Zotero extract structured metadata automatically
• GitHub Pages docs at giovannipicogna.github.io/lmu-usm-agent-template

• agents page — all 12 agents with example invocations
• skills page — bundled skills by domain
• Auto-deployed on every push to main via GitHub Actions

Settings → Pages → Source: GitHub Actions

copilot-instructions.md §11: domain-specific skill lists for every research area

• Embargoed or proprietary observational data (pre-publication)
• Unpublished simulation outputs (IP of your group)
• Patient / medical data — EU GDPR compliance
• HPC login nodes without outbound internet access
• Budget constraints — zero API cost once hardware is available
• Offline development or conference travel

111 tasks · 20 peer-reviewed astrophysics papers · tasks co-written by original authors

success rate for best frontier models

• Tasks include: Abacus N-body near-field summation with spline softening; B-spline localisation of black-hole merger-rate peaks in mass
• Failure mode: coding is fine — domain-specific rigor is not

120 tasks · 100 synthetic + 20 real RV datasets · multi-planet systems

• Agents achieve optimal statistical fits on periodograms — but fail to recover true Keplerian elements
• Difficulty scales with: SNR, orbital resonances, planet multiplicity
• Root cause: no persistent physical world model — statistical optimisation ≠ physical understanding

• Task: ACT DR6 CMB power-spectrum analysis
• PI + Student agent hierarchy
• Shared "Blackboard" for async cross-agent memory retention
• Open-ended: hypothesis → computation → diagnostic iteration
• Closest analogue to our @pipeline-agent design

• Task: out-of-distribution detection in weak-lensing maps
• Structured tree-search + dedicated idea sampler
• Iteratively refines detection algorithms against known physical bounds
• Excels where objectives are mathematically precise and measurable

• Task: end-to-end paper generation
• Autonomously applies quantum-physics methods to astrophysical datasets
• Expert-validated papers in astrophysics, planetary science, mathematical physics
• Still requires domain expert review — not autonomous publishing

+26.7 percentage points vs baseline — zero fine-tuning required

• 199 domain-specific SKILL.md files (biostatistics, data analysis, scientific writing)
• Agent reads structured Markdown → applies domain knowledge dynamically at inference time
• Directly validates our dustpy, pluto, fargo3d, radmc3d, sherpa, yt skills

Science question:
"What mechanisms set the depth of a planet-carved gap in a protoplanetary disc? Vary M_p, α, and v_frag."

Agent ran 3-round internal debate → cross-checked NASA ADS

Hypothesis: "1 M_Jup planet
opens deep gap in M★=1 M⊙ disc
with h/r=0.05, α=1e-3."
Compute gap-opening criteria
and Kanagawa gap depth.

Find the key papers on the fragmentation barrier in dust coagulation and export BibTeX to paper/bibliography.bib

• 6× ads_search queries
• ads_resolve_reference for 2 ambiguous refs
• ads_citation_helper on 7-paper core set
• ads_export batch → BibTeX, 1 call for all 12

Generate FARGO3D par file for
gap depth run: 1 M_Jup at 20 au,
M★=1 M⊙, h/r=0.05, α=1e-3,
Nx=384×Ny=128.
Add SLURM script for LRZ.

python .github/skills/dustpy/\
  scripts/run_dustpy.py \
  --alpha 1e-3 \
  --disk-mass 0.05 \
  --stellar-mass 1.0 \
  --r-in 1 --r-out 100 \
  --N-r 80 \
  --vfrag 10 \
  --t-end 1e6 \
  --N-snaps 10 \
  --output-dir data/dust/iceline_test

python .github/skills/fargo3d/\
  scripts/run_fargo3d.py \
  --par-file setups/fargo/fargo.par\
  --output-dir data/runs/fargo_test\
  --AspectRatio 0.05 \
  --Sigma0 6.37e-4 \
  --PlanetMass 9.548e-4 \
  --Tmax 5 --Ninterm 5 \
  --fargo3d-bin fargo3d \
  --n-procs 1

python .github/skills/pluto/\
  scripts/compile_pluto.py \
  --run-dir .../Test_Problems/\
             HD/Disk_Planet \
  --pluto-dir .../pluto-code \
  --with-fargo \
  --make-jobs 4

import h5py, numpy as np
import matplotlib.pyplot as plt

for snap in [0, 4, 9]:
  f = h5py.File(f"data{snap:04d}.hdf5")
  r = f["grid/r"][:] / 1.496e13  # au
  a = f["dust/a"][:]   # (Nr, Na)
  Sd = f["dust/Sigma"][:]
  St = f["dust/St"][:]
  i  = np.argmax(Sd, axis=1)
  a_max = a[np.arange(Nr), i]
  St_max = St[np.arange(Nr), i]
  ax[0].semilogy(r, a_max)
  ax[1].semilogy(r, St_max)

Post-process the FARGO3D output
in fargo_skill_test/:
  - Read gasdens0.dat (128×384)
  - Azimuthal-average → Σ(r)
  - Compute Σ_gap / Σ₀
  - Read tqwk0.dat (planet torque)
  - Save JSON + radial profile plot

Interpret analytical +
numerical results:
K=2916, Σ_gap/Σ₀=8.5e-3 (theory)
vs Σ_gap/Σ₀=1.0 (simulation t=0)
DustPy: a_max(100au)=24cm
Verdict: iterate or stop?

Draft abstract + methods for
the disc gap study.
Use ONLY ADS-verified bibcodes.
Mark uncertain refs [VERIFY ADS].
Save as LaTeX to
paper/sections/abstract_methods.tex

9 agents · 3 skills · 13 ADS-verified references · science question → draft methods in one session

Gate 1: resource & direction decision. Gate 2: validity & cost decision. Agents propose — humans decide.