Agentor Agent Framework

The README says tracing is off by default; when enabled, it includes prompts, tool arguments, tool results, and provider-returned reasoning, and can be disabled per run. This supports limited data-flow transparency (2). However, the API-key example reads from an environment variable, the authentication example relies on users checking a token themselves, and no global permission model, sensitive-data protection, or user confirmation for destructive tools is shown, so least privilege, confirmation, sensitive-data handling, and control of external effects score 1. Persistent event logs support resume and fork (rollback 2), but deletion, recovery failures, and compensation for external side effects are not explained. Dependencies have lower bounds and optional extras (1), but there is no evidence of locking, auditing, or vulnerability mitigation. LICENSE, author metadata, and a vulnerability-reporting process provide limited attribution; the publisher is unverified and no clear maintainer commitment is given (source_attribution 1).

The README describes durable runs, resume, fork, and sync/async interfaces; the test workflow covers three operating systems and Python 3.11–3.13, showing some engineering support (1, 2). However, the README describes the v0.1.0 line while the test's expected agent card contains version 0.0.1, an inconsistency between documentation and test expectations that lowers self-consistency. Comments say missing optional dependencies produce a clear ImportError, but there is little evidence of failure paths or user-facing error examples, so failure_messages scores 1.

The developer-facing documentation covers simple agents, ASGI serving, MCP, A2A, skills, durable runs, and several LLM endpoints, showing multiple scenarios and environments (audience_and_scenarios 2, environment_fit 2). But safety and capability boundaries are mostly feature claims; tool permission rules and suitable or unsuitable use cases are not clearly defined. Triggering is illustrated through explicit Python calls, a serve entry point, and per-run tracing options (trigger_precision 2), but boundaries for automatic triggers and side-effecting tools are not explained.

The README is organized by installation, features, and examples; pyproject provides package metadata, a minimum Python version, and dependency extras, with reasonably clear optional-dependency installation instructions (information_architecture 2, install_notes 2). Agentor naming is stable at repository level, but the product description, package version line, and default agent-card version in tests do not fully align (naming_stability 1). Examples cover several features, and the security policy specifies reporting and supported versions; no FAQ is shown (examples_and_faq 2). The policy says older tags are unsupported and response targets are non-binding (known_limitations 2). The Apache-2.0 text is complete and consistent with metadata (license 3). There is a tagged release workflow and dynamic version source, but no changelog is provided (versioning_changelog 1). The security policy describes a small team, best-effort handling, and a private vulnerability-reporting path (maintenance_responsibility 2), but identity is not verified by the enterprise registry and there is no binding maintenance commitment.

The examples provide practical starting points for a weather agent, serving requests, durable runs, Markdown-defined agents, MCP, and A2A (output_usability 2). Durable runs, tool search, MCP integration, and A2A suggest value beyond basic model calls (marginal_value 2), but key superiority claims such as “fastest” and “secure” lack comparison methods or quantitative evidence; the core dependency set is substantial and cost-benefit discussion is limited (cost_benefit 1).

Some README features correspond to configuration, CI workflows, and tests; for example, tests cover the A2A card and import timing, but this does not establish that features were executed or achieved their advertised results (claim_traceability 1). README, pyproject, LICENSE, security policy, and workflows provide limited corroboration about installation, licensing, publishing, and reporting, but the core product features have little test corroboration (1). The documentation presents code examples, feature descriptions, and claims such as “fastest” and “only” as distinguishable kinds of content, allowing visible facts to be separated from unverified performance or safety inferences (2).