Execution hooks for programmatic tool invocations

[nayname]

What would you like to be added?

Context

When using ElcAgent programmatically (e.g., in automated pipelines), the LLM controls the execution loop. The calling application has no visibility into the intended execution path, and no way to review or approve the plan before tools start running.

Why this matters

In non-interactive contexts (CI/CD, backend services, scheduled jobs):

• Tool calls are initiated by the model
• Parameters are model-generated
• Execution happens immediately once a tool call is emitted

The execution path is implicit and non-previewable, making it harder to reason about
intent, correctness, and side effects.

Proposal

An execution layer between user intent and tool execution.

Rather than letting the LLM drive execution directly, the system would:

1. Receive user intent
2. Generate a deterministic execution plan (potentially multi-step)
3. Allow the calling application to review/approve before any execution
4. Control workflow execution externally to the LLM

Key concepts:

• Execution Plan — user intent produces a deterministic, reviewable plan before
  execution. Plans can be multi-step and are validated before any tool runs.
• Guardrails — declarative rules (parameter bounds, allowed operations, environment
  restrictions) enforced before execution.
• Preview → Approve → Execute — the calling application controls when execution
  happens, not just whether it's allowed.

Core principle: the LLM proposes, infrastructure enforces.

Integration

I've been working on this pattern in a separate project (IAMY in my profile). Happy to:

• Share it for discussion
• Submit a PR if there's interest
• Collaborate on what pattern fits best (config option, wrapper, event emitter)

Is this direction interesting for the project?