Coaching and Corrections

Coaching lets 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. The agent's model weights remain fixed during generation.

Who does what?

Component	Responsibility
Scenario and executor	Define and observe success, failure, budgets, and tool results
GEN	Select opportunities, assemble evidence, choose allowed restart points, and evaluate continuations
Agent package	Define its repair strategy and render corrected decisions with consistent memory
Configured LLM backends	Assist automatic diagnosis and agent-specific repair
Human in the viewer	Select an executed decision, supply a diagnosis, and propose its replacement

Diagnosis and repair

A failed action does not necessarily identify the decision that caused the failure. The agent may have used a wrong argument, missed a prerequisite, retained an incorrect assumption, or encountered an injected environment error. A valid physical action can also fail during execution. Useful supervision must connect the observed context, the proposed correction, and its consequences.

Automatic coaching can start after a failed stage, invalid format, exhausted budget, lack of useful action, or suboptimal completed trajectory. Routing, agent capabilities, and collection budgets determine which opportunities are eligible; each unsuccessful tool call does not necessarily trigger a correction.

GEN supplies stage objectives, decisions, feedback, and relevant error information. Diagnosis identifies a decision worth revisiting and explains the suspected cause. For a trajectory with only one diagnosable decision, diagnosis can be deterministic. Format repair and text rewriting have their own routes; the process does not require three LLM calls for every correction. A suboptimal diagnosis can decline to propose a repair.

The agent owns the strategy appropriate to its architecture. It might replace an action, repair a summary, or revise a plan. GEN sends the resulting correction through the normal agent response endpoint so the new decision and memory stay consistent before execution resumes.

Counterfactual re-execution

A replacement tests what happens if the agent makes a different decision from this saved point. GEN preserves the original branch and creates an alternative using the saved input, simulator state, and execution context.

Saved situation S
  ├─ original decision → original outcome
  └─ diagnosis + intervention
       → agent renders corrected decision and memory
       → execution from S → new feedback and task checks

For example, an agent tries to place an object before picking it up. A diagnosis can select that decision and propose picking up the object first. Re-execution checks the pick from the saved situation; the continuation still has to place the object and satisfy the task's conditions.

An automatic failure route can also offer recovery after suitable invalid call feedback. This retains the failure feedback and asks what to do next. Restart anchors distinguish recovery from replacement of an earlier decision.

Human corrections in the viewer

It is also possible to a human to propose correction directly through the viewer. Take a look at the viewer guide.

Validation and training evidence

GEN checks the new continuation using tool feedback, physical/log goals, applicable semantic validation, and execution budgets. The new branch can succeed, remain ongoing, fail again, or stop at an exploration limit.

Outcome	What it establishes
Proposal produced or correction accepted	An intervention is available for reinjection
Candidate rendered and executed	The intervention became a real continuation
Successful evaluated continuation	The relevant task checks were satisfied

A coached marker identifies origin. A successful local correction does not certify the rest of the task, and acceptance does not guarantee a training row. The exporter selects candidates using observed downstream outcomes. Validated examples can also be registered for later automatic coaching; that cache reuse does not train the agent during a run.

Capabilities and budgets

GEN uses the support announced by the agent. Specialized correction kinds are format, failure, and suboptimal; generic text-resume support is advertised separately. Agents may implement only some routes. Adding a new internal model or memory component does not automatically provide a suitable repair strategy.

GEN owns backend configuration and supplies effective settings to supported agent coaching sessions. Those sessions use LLM backends and optional log capture. Human interventions use viewer controls rather than a separate provider. Inference and coaching share Core's HTTP protocol 2.0. Returned logs are saved by GEN; the agent does not need the output directory mounted locally.

Automatic work has per-stage pending and unproven limits, both defaulting to five. Pending includes queued, running, and answered requests awaiting reinjection. Unproven work counts pending plus completed attempts minus attempts that led to stage success. Human interventions are exempt from those two limits, while normal stage/exploration budgets still apply.

To enable assistance and reuse examples, follow Configure Coaching.