Packages

beam_agent

0.1.0
0.1.0

A general-purpose, OTP-based agent harness for Elixir: an LLM tool-use loop with an injectable tool registry, configurable guardrails, pluggable post-run verification, deterministic context compression, a full execution trace, and supervised, concurrent agent runs.

HexDocs Report package

Current section

Files

Jump to
Readme
1 Version
0 Dependencies
Dependants
Files
Activity
beam_agent CHANGELOG.md
Raw

CHANGELOG.md 

# Changelog



All notable changes to this project are documented in this file.



The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),

and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).



## [0.1.0] - 2026-08-27



Initial release.



### Added



- `BeamAgent.API.run/2` — the public entry point: runs a goal to completion

  under supervision and returns `{:ok, run}` / `{:error, run}`.

- `BeamAgent.Runner` — the LLM-driven tool-use loop, as an isolated,

  supervised `GenServer` per run (`BeamAgent.RunSupervisor`), so any number

  of runs can execute concurrently and one crashing has no effect on the

  others.

- `BeamAgent.Context` — deterministic (no LLM call) context compression

  that keeps the model-facing message count bounded as a run goes on.

- `BeamAgent.Guardrails` — independently configurable `MaxIterations`,

  `MaxContextMessages`, `MaxExecutionTime`, and `MaxToolCalls` checks.

- `BeamAgent.Verifier` — pluggable post-run verification

  (`BeamAgent.Verifier.Behaviour`, with `BeamAgent.Verifier.Default`

  checking execution status, answer presence, and required tools) that

  catches a model claiming success without doing the work.

- `BeamAgent.LLM.Client` — the one-callback behaviour a real LLM provider

  implements to plug into a run.

- `BeamAgent.Tools.Behaviour` / `BeamAgent.Tools.Registry` — an injectable

  tool registry; each run supplies its own `tools` map rather than the

  harness shipping a hardcoded one.

- `BeamAgent.Run` / `BeamAgent.Trace` — the structured result and full,

  inspectable execution trace returned on every run, success or failure.