Coaching and Corrections

Coaching helps MAGMA-GEN explore a corrected continuation when an agent produces a malformed decision, fails a task, or follows a suboptimal trajectory. It can make recovery examples available for later training. It does not update the agent's model weights during the run.

Who does what?

Component	Responsibility
Scenario and executor	Define and observe success, failure, budgets, and tool results
GEN	Select correction opportunities, assemble context, choose allowed restart points, and evaluate new continuations
Agent package	Define its own correction logic and how corrections affect internal models, memory, and decisions
Configured coaching provider	Supply LLM or human assistance to the correction process

A verifier deciding whether an answer is acceptable and a coach proposing a better continuation have different roles. The same external model service may support several roles without becoming the agent under evaluation.

Correction lifecycle

Agent decision → execution / validation → correction opportunity
  → diagnosis and agent-specific repair
  → proposal at an allowed restart point
  → agent processes the corrected input
  → execution / validation of the new continuation

A proposal is not proof of success. GEN sends the corrected input through the agent's inference interface so the agent can produce a coherent decision and memory, then checks the new execution. Earlier records remain available along with correction provenance.

Examples include repairing invalid output syntax, replacing a failed action, or revising a plan or memory that caused a poor decision. Each agent decides what can be repaired and how. A summarizing agent might need to correct its summary and rerun its decision model; another agent may replace only one action.

Optional capabilities

GEN uses the support advertised by the agent. Specialized correction kinds are format, failure, and suboptimal; support for resuming generic text corrections is advertised separately. Do not assume every agent implements every correction route.

Full-history currently implements specialized corrections and replacement decisions. The history-summarized example currently leaves corrections disabled; adding a summarizer requires an appropriate repair strategy, not automatic reuse of full-history's memory edits.

GEN owns the coaching provider configuration and sends effective settings to supported agent coaching sessions. Agent inference/model settings stay in the agent package's configuration. Logs return to GEN for recording; the agent does not need the run's output directory mounted locally.

For operating GEN and interpreting tested interventions, read Diagnose, propose, validate.