LoginSign Up
← Back to all sparks
M

mini007

AI-ASSISTANTS
Velocity0.0

Lightweight Framework for Orchestrating Multi-Agent Large Language Models

github.com/feddelegrand7/mini007 ↗

mini007 gave its R agents tools and a way to argue with each other.

llm-agentstool-callingmulti-agentellmerr6r-package
◆Current state
mini007 is an R multi-agent framework built on R6 classes over ellmer, with a LeadAgent that generates a plan and delegates to sub-agents. Over eight months it went from conversation plumbing to a working agentic surface: message history as a mutable active field, budget limits and policies, in-session R code generation and execution, plan visualization, and — from 0.3.0 — tool registration and a two-agent dialog mode.
◆Where it's heading
The package is assembling the standard agent-framework feature set in the order most frameworks reach it: memory management first, then cost control, then tool use, then multi-agent interaction. It is tightly coupled to ellmer, which it took on as a hard import in 0.2.2 after sync problems, so its ceiling is set by what ellmer exposes. The release record is thin and imprecise — the 0.4.0 notes are a verbatim copy of 0.3.0's, so whatever actually shipped in May 2026 is undocumented.
◆Prediction
Given the trajectory from two-agent dialog, the next step is most likely more agents in a single conversation or richer delegation topologies. The duplicated release notes make it hard to say what is already in progress.

◆Recent moves

  1. 2mo ago

    Release notes repeat 0.3.0 verbatim

    The 0.4.0 notes are byte-for-byte identical to 0.3.0's, re-listing the tool methods and agents_dialog() that shipped four months earlier. Whatever changed in this release is not recorded, so there is nothing here to assess.

    View source ↗
  2. 7mo ago

    Tool registration and two-agent dialog arrive

    ⚡ SPARK

    The release that turns mini007 from a conversation orchestrator into an agent framework proper. Everything before it manages what an agent remembers and spends; this is the first release that changes what an agent can do.

    View source ↗
  3. 8mo ago

    ellmer becomes a hard import; response validation added

    ellmer moves from loose coupling to a declared import dependency after synchronization problems between the two, and validate_response() and clone_agent() join the API. Taking the dependency explicitly is the tradeoff that makes the later tool work possible.

    View source ↗
  4. 9mo ago

    Bug fix in generate_execute_r_code()

    A single unspecified fix to the code-execution method introduced days earlier. Follow-up patch with no described user-facing change.

    View source ↗
  5. 9mo ago

    Message history, budgets and code execution land together

    Eleven methods arrive at once, covering the three things a long-running agent needs: memory control (keep_last_n_messages, clear_and_summarise_messages, export and load history), cost control (set_budget, set_budget_policy), and action (generate_execute_r_code, visualize_plan). Making messages an active R6 field is the structural change underneath. Plan regeneration becomes opt-in via force_regenerate_plan, and delegate_prompt() is dropped as redundant.

    View source ↗