Why This Field Matters

Coding agents have moved past five-minute autocomplete. In 2026, OpenAI published a stress test where Codex ran for roughly 25 hours uninterrupted, burned about 13 million tokens, and produced nearly 30,000 lines of code. This isn’t a one-shot prompt anymore — a manager agent decomposes the work, dispatches it to parallel workers, and the system spends hours building, testing, and fixing on its own. This is the era of long-horizon autonomous work. METR’s time-horizon metric estimates that the length of task a frontier model can handle at 50% reliability has roughly doubled every seven months since 2019. If that curve holds, day-scale tasks land around 2028.

But the longer a model runs, the more the real bottleneck shifts from the model to the harness wrapped around it. Anthropic dedicated an entire engineering piece to effective harnesses for long-running agents, because the infrastructure that maintains state, mediates tool calls, verifies progress, and catches drift is what governs reliability. If a five-minute call fails, you just call it again — but if an agent that has run for eight hours collapses at the end, all eight hours are gone. Saving intermediate state with checkpoints, pausing and resuming across process boundaries, capping token budgets so cost doesn’t run away, and pausing safely where human approval is required — the person who designs all of this is the long-horizon agent orchestration engineer.

Required Skills

This work sits at the intersection of distributed-systems engineering and AI engineering. Think of it as reliability engineering for non-deterministic workers. On top of that, long-running execution adds its own instincts.

• Harness and orchestration design. Build manager-worker structures, split context windows per subagent, and handle task decomposition and reassembly. The longer an agent runs, the more context leaks, so you deliberately design what it should remember versus compress and discard.
• Checkpointing and resumption. Build execution that survives stop-and-restart. Save intermediate state, retry from the point of failure, and write durable workflows that survive a deploy happening mid-run. Idempotency is the baseline.
• Drift and reliability guards. Stop an agent from wandering off its original goal hours in, or looping on the same mistake forever. Progress verification, loop detection, per-step gates, automatic rollback — you wrap deterministic safety nets around a non-deterministic system.
• Cost control and observability. Cap token budgets and watch unit cost in real time. Instrument calls, tool use, and reasoning against OpenTelemetry’s GenAI conventions so you can read back where an eight-hour run spent its tokens.
• Human-in-the-loop resumption. Design the points where a human steps in. Build flows that pause safely and resume without losing context after a person has reviewed.

Career Path

Demand is climbing steeply, yet few people have actually shipped a long-running agent to production. So this role asks for an awkward intersection — not a generic backend engineer, not a pure ML researcher. The center of gravity is the mid-to-senior engineer who knows distributed systems and reliability engineering and has also wrestled with the non-determinism of agent runtimes. Can you solve “the demo runs but production collapses after 30 minutes”? That’s the question.

The way in is surprisingly ordinary. Start in SRE, platform, or distributed-systems backend and take on agent infrastructure, or come down from AI engineering’s orchestration side into the reliability and runtime layer. Titles haven’t settled yet, so the work scatters across Agent Infrastructure Engineer, AI Platform Engineer, and Agent Reliability Engineer. Whether you’re at a FAANG company or a Series A startup, running internal dev agents reliably is no longer a side project — it’s the platform team’s actual job. Compensation tracks the top of the AI infrastructure and platform band, often the upper end of the platform or infra track.

The fastest way to prove it is to build one. Wrap a single coding agent in a small harness, make it stop and resume from a checkpoint, add a token budget and loop detection, and instrument every call with OTel. Then kill it mid-run on purpose and measure whether it comes back cleanly. That one cycle beats any keyword on a resume.