Use when each evaluation in a black-box search is costly (a training run, a simulation, a benchmark), you hold a trace of configurations already evaluated with their scores, and you are about to spend more budget on blind trial and error; it fits a checked…
Use when a task should be worked by a team of AutoGen agents (taking turns, picked by a model, or handing the task over) on a model the person passes in, with a message limit G6's trusted policy caps, every tool call decided by G6's gate, G6's stop checked…
Use when a multi-step workflow should run as a LangGraph graph that saves its state after every step under G6's state folder, so a run that is stopped, killed or fails resumes where it was without repeating finished steps, with every tool call decided by G6's…
Use when a task suits a smolagents agent, either calling declared tools or writing Python to work it out, on a model the person passes in, with every tool call decided by G6's gate and every code step held to a restricted syntax and run in a bounded child…
Use when an answer, a test expectation or code rests on a symbolic calculation (simplifying, factoring, solving an equation, a derivative, an integral, a limit, a series, a matrix's determinant, inverse or eigenvalues, or an integer's factors, primality or…
Use when an answer says one thing causes, drives or improves another, or estimates an effect from data, and you are about to choose which variables to control for, call an effect identifiable, or report a causal number from a correlation, a regression or your…
Use when code has a fixed shape with a few unknown parts (constants, a sign, a comparison, a choice between strings) and you are about to pick those parts by hand from a handful of examples, or to trust generated code because it passed the examples it was…
Use when a coding agent is about to act on a risky plan (a deploy, auth or security, credentials, a data migration, payment, medical or legal logic, skipping or disabling a check) or is about to report coding work as done, and the plan or the drafted answer…
Use when an agent's plan could reach a harmful state and nobody has worked out how likely that is: write the agent as a small model of states and transition probabilities with the unsafe states named, then get an upper bound on reaching one within N steps,…
Use when an answer depends on what follows from a set of facts and rules (permissions, eligibility, dependencies, reachability, policy, family or ownership relations) and you are about to say that something does or does not follow, or to list everything that…
Use when an answer rests on a logical step that must be exactly right, such as whether premises entail a conclusion, whether constraints can all hold, a counterexample to a claim, two conditions being equivalent, which requirements conflict, or a small…
Use when you need to read or change data in a database the person running G6 has declared (SQLite by default, DuckDB for large datasets or files, PostgreSQL), and want reads that cannot write, changes dry-run first in one transaction, destructive statements…
Use when a result envelope (from a G6 component, your own tool or another agent) is about to be trusted, passed on or asserted in a test, and you need to know whether its completion_state, warning_card, evidence, request_id and run_id agree with one another,…
Use when an agent works with people who may be in distress (a support chat, a care or clinical intake, a wellbeing check, a carer's notes) and must not miss signs that a vulnerable person is at risk of suicide, self-harm, harming someone else or a child being…
Use when an agent must rank a set of passages it holds by meaning rather than shared words (a question worded differently from the text that answers it), and must know which method ranked them, so word matching never stands in silently for meaning, model…
Use when a PDF, a web page, Markdown or a text file must be read as text an agent can cite, and it matters what was not read, so a page without a text layer, hidden page text or a format nothing converts is reported rather than silently dropped or read as…
Use when you need to cut, join, mix, fade, level, equalise, compress, limit, gate, de-ess, de-noise, resample or convert a WAV, or measure its loudness, true peak, pitch, tempo, spectrum or silences, and you want every result measured again before you use it,…
Use when you need a 3D scene built, lit, rendered or exported (OBJ, STL, GLB, FBX) with the Blender the person installed, and you want every change read back from Blender's scene and every render and export inspected before you use it.
Use when you need a picture made or edited from a text prompt by a diffusion model on the ComfyUI server the person runs on this machine (text to image, image to image, inpainting, upscaling, or your own ComfyUI workflow), with every picture fetched into G6's…
Use when you need a diagram (a flowchart, a sequence of messages, a graph, a component map, a box, tree or table in plain text) written as Mermaid, Graphviz DOT, PlantUML or ASCII that will not break on a quote or a keyword, or drawn by Graphviz into an SVG…
Use when you need an actual picture file and not a description of one (procedural art, a texture, a pattern, a fractal, a vector drawing or a terrain mesh for a 3D scene), made on this machine with no model, no network and no cost, from a seed you can…
Use when you need to resize, crop, rotate, flip, recolour, combine or convert a picture, measure one strictly, or have one generated (locally through ComfyUI or a procedural technique, or by a paid image service only when G6's action gate allows the payment),…
Use when you need a tone, a note, a chord, an arpeggio, a scale, a short melody, noise or a drum pattern as a WAV file, or want a few words read into musical parameters, and you want each sound measured against what you asked for before you use it.
Use when you need a chart of rows of data (a line, bars, a scatter, a histogram, box plots or a heat map) as an SVG, PNG or self-contained HTML file, and need to know it shows every row you gave and only the values you gave, with nothing dropped silently and…
Use when a repeated task fails, costs too much or runs slowly, and you have labelled development and held-out cases for testing proposed changes before a person activates them.
Use when you want the best of several scikit-learn estimators and their settings for a table, found by a bounded, seeded search (a time budget and a trial cap) that compares settings by cross-validation on training rows only, with the winner measured on rows…
Use when a scalar beta-binomial, gamma-Poisson or normal mean model needs an explicit posterior, credible interval, prior sensitivity, predictive check or checked sampling, and a point estimate would hide uncertainty.
Use when you need a small neural network trained on rows of data without writing the training loop yourself, and need its score measured on rows held out by the tool itself, against a baseline, with the run seeded and stopped the moment its loss breaks, so a…
Use when you need to fit a scikit-learn model to rows of data, or compare a few, and need its score measured on rows held out by the tool itself, against a baseline, with an interval, after checks for target leakage, repeated rows, class imbalance and a fixed…
Use when a design has states and steps (a protocol, a lock, a retry loop, a workflow, a state machine, concurrent or distributed steps) and you are about to say it never reaches a bad state, never deadlocks or can always reach a goal, from tests or a reading…
Use when you are about to propose a schedule, allocation, assignment, route, flow, packing or production plan that must keep stated rules, call a plan the best, cheapest or fastest, or tune a non-linear formula's inputs by trial and error, and the numbers…
Use when an answer depends on what a Prolog program derives (rules with recursion, lists, arithmetic or terms with structure) and you are about to state the answers, or that a goal does or does not follow, from your own reading of the program instead of…
Use when finishing a Python coding task (review the patch with G6 before saying it is done), constructing AI applications from checkable components, investigating repeated failures, testing redesign hypotheses, or reviewing claims and risky actions against…
Use when an agent must see what is in a picture (describe it, answer a question about it, read its text, list its objects) through an image-capable model it was given, and must keep the model's account as a claim, with the picture checked before it is sent…
Use when a coding agent is about to fix a failing test or change a function in a local git repository and what counts as done must be frozen and confirmed by a person before the patch is written, so the patch is checked against that standard rather than…
Use when an agent proposes an outward action and needs a deterministic decision before acting, when a workflow must stay inside the bounds a person reviewed it for, or when the person wants per-run operation count limits and injected safeguards that can only…
Use when an agent or workflow is about to take a large or hard-to-undo step and must check that it has a complete plan to observe, stop and roll it back, that it does not collide with other actions in flight, that no warning signal is being ignored, or that…