← Back to feed

Developers using GPT-5.6-Sol in multi-agent orchestration pipelines are hitting a consistent failure: the model ignores explicit instructions to terminate sub-agents at phase boundaries, causing runaway token consumption. One developer reported over $500 burned in a single Codex session while following a documented, phase-gated plan.

The failure has two components. First, the orchestrator does not close sub-agents even when AGENTS.md explicitly instructs it to start fresh at each phase change. Sol continues delegating to persistent sub-agents that accumulate context from previous phases. Second, the model re-runs the full test suite at every intermediate code change rather than waiting for the review-ready checkpoint specified in the plan.

What Fails and Where

The failure is at the orchestration layer, not at base capability. GPT-5.6-Sol at very high effort processes its reasoning expansively and loses track of operational constraints defined in system documents. The lifecycle instruction — “start from a fresh sub-agent when changing phase” — is treated as a soft preference rather than a hard rule.

The result: instead of spawning a contained sub-agent for a bounded task, the orchestrator runs token-heavy inference loops that accumulate context and compound cost across phases.

This is distinct from the July 2026 Codex subscription cap issue, where OpenAI’s billing throttle was burning users’ $200/month Pro allocations. That problem was addressed with an 18% efficiency improvement in a July 29 patch. The August incidents are about application-layer orchestration behavior — the model’s runtime decisions about sub-agent lifecycle — not billing cap mechanics.

Community Workaround

The effective mitigation documented by developers building production agents on Sol:

1. Define the editable surface explicitly. A project-specific document scopes exactly which files and directories the agent may modify. Without this, Sol treats the entire codebase as in-play and over-explores.
2. Constitutional prompt design. The system prompt encodes style guides, design principles, and explicit boundary conditions as first-principles. Sol responds better to constraints framed as fundamental rules than as operational instructions buried in markdown.
3. Hard phase gates in the orchestration harness. Do not rely on the model to self-terminate sub-agents. The calling harness should enforce phase transitions and kill lingering contexts explicitly.

The pattern is familiar to anyone who has managed a highly capable but poorly scoped junior engineer: the model performs well within a defined surface, but left to self-organize, it over-explores and over-executes.

Cost Exposure

At very high effort, GPT-5.6-Sol’s compute cost is high. An orchestration failure that keeps sub-agents running across multiple phases can reach $200–$500 in a single session. At production scale — teams running dozens of parallel agent tasks — a single misconfigured pipeline can consume a month’s API budget in hours.

The problem is upstream of any individual efficiency fix: multi-agent orchestration at the frontier lacks the operational tooling (lifecycle hooks, context scope limits, cost circuit breakers) that would make these systems production-safe without heroic prompt engineering. The model is capable. The harness around it is not.