What survives from long-running agent harnesses when models get strong (mid-2026 postmortem)

A developer who built a long-running pipeline harness for coding agents found that frontier models (Fable 5, GPT-5.6 class) and platform vendors made the pipeline itself unnecessary. The surviving value is a set of operating principles: never-self-verify, evidence over assertion, isolation scaling with autonomy, orthogonality for parallelism, stateless outer loop with single writer, and doc quality equals throughput.

I built and shipped a full long-running pipeline harness for coding agents spec → orthogonal feature decomposition → parallel TDD subagents in git worktrees → stateless reconciler → human verify gate → topological merge , ran it on a real 482-commit ERP project, then repackaged it as a cross-agent skill Claude Code / Codex / Gemini . Weeks later, frontier models Fable 5, GPT-5.6 class made the pipeline itself unnecessary: they decompose and sustain long-running work natively, and the harness vendors absorbed the orchestration layer native subagent fan-out, worktrees, sandboxing, workflow primitives . The lesson: a harness does two jobs. 1 Compensating for weak long-horizon coherence — checklists as external memory, task-slicing, babysitting loops. This half is obsolete: the model ate it. 2 Coordination, trust, safety, and parallelism — this half was never about model strength, but it got absorbed by the platforms instead. What survives is not a pipeline. It is a paragraph of operating principles: Never-self-verify. The agent that did the work never flips its own status to "verified". "I did it" and "it's been verified" are different trust levels, and the boundary between them is where autonomous systems earn or lose trust. Evidence over assertion. Verification rests on artifacts a human can inspect — screenshots, traces, transcripts — bound to the exact commit SHA they claim to prove. This matters more with strong models, not less: a stronger model games proxies tests, "it passed" more convincingly. Where automation can't reach, say so honestly instead of faking a seam. Isolation scales with autonomy. The stronger the model and the longer it runs unattended, the harder the boundary it needs: throwaway $HOME, container/microVM, default-deny egress. Permission prompts are not an isolation strategy. Orthogonality is the price of parallelism. Parallel agents need disjoint file ownership. Kill shared entry files routers, DI containers, init with codegen registries; give merge-ordered artifacts migrations, event logs timestamp names, not sequence numbers. This is a property of parallelism itself — no model strength removes it. Stateless outer loop, single writer. Any coordinating loop re-reads all state from disk every iteration; exactly one writer owns the source-of-truth file. Everything else reports through mailboxes. Interruptible, resumable, human-editable. Doc quality = throughput. Anything fed to a context-free worker spec, plan, checklist gets reviewed before it ships downstream. An ambiguity you'd resolve in conversation becomes a coin-flip in a worker with no conversation to ask. Build the principles into your instructions file. Let the model and the platform own the machinery.