What is an agent harness?

Anthropic's S-1 defines it as the secure environment, credentials, memory, and orchestration loop a model needs to plan, call tools, verify its work, and complete a task safely on its own, with the summary line that a model alone is not an agent. The filing offers it two ways: the Agent SDK, which makes the same production-grade harness that powers Claude Code available to run on a customer's own compute, and Claude Managed Agents, where Anthropic hosts the sandboxes, stores the credentials, maintains the memory, and runs the agent loop while customers define tasks, tools, and guardrails.

What does a forward deployed engineer build now that the harness is a product?

Connectors, skills, evals, and governance. The S-1 says intelligence is one limiting factor but so are connectors that do not yet exist, governance frameworks that are not yet written, and orchestration logic that needs to be built from scratch, and that custom skills capture anything from engineering standards and regulatory processes to proprietary toolchains. Colin Jarvis's account of OpenAI's team has the same list: when a rebuilt deployment reached the next team, the work was swapping connectors and building fresh evals, for which the team had a skill. The handback to product is a skill file and a row set rather than a repository.

Why is week one an audit rather than an install?

Because the tool arrives before the engineer. The S-1 says that by the time Anthropic's direct teams engage a customer, Claude is often already in use, and that in the quarter ended September 30, 2026, Claude Enterprise users without Claude Code spent less than $75 a month on average while the top 5 percent with it spent about 20 times that. Rebecca Bellan at TechCrunch reported that more than half of customers spending $100,000 or more a year use both the platform and a first-party app, up from about 30 percent. So the FDE lands where developers already run the harness, with skills, connectors, and prompts nobody versioned, and the first job is to list what is running, who wrote it, when it last changed, and what checked the change.

What did the tooling not change for the FDE?

Verification. Lesson 08 of Becoming Forward Deployed Engineer, published October 26, 2026, says a prompt is code with nothing checking it: four lines trimmed on a Thursday to save tokens, complaints from the adjusters about two weeks later, and no way to connect the two. Its answer is a prompt registry that stores each version with the eval score written by the same function that ran the eval. Anthropic's S-1 says over 90 percent of contributions to its own codebase in September 2026 were written by Claude and that code volume per engineer rose over 45 times since 2024. Changes got cheaper. Checking did not, so the eval on every change, every model swap, and every prompt edit is the FDE's job.

What is the Claude Frontier Academy and why does it matter for the FDE shortage?

The S-1 says Anthropic launched the Academy in October 2026 with a $100 million commitment to address the shortage of engineers who can take AI from an initial use case into production, training customers' and partners' engineers to the same standard as its own, with a goal of 10,000 engineers by the end of 2027. Alongside it the filing describes Ode with Anthropic, an AI services company for mid-sized firms that will partner with the Applied AI teams. The vendor's answer to the shortage is to turn the seat into a product, a training program, and a services firm, which leaves the FDE the part that was always the point: the eval and the handback.

What Agent Harnesses Changed for Forward Deployed Engineers

What did agent harnesses change for forward deployed engineers?

THE SHORT ANSWER

They moved the forward deployed engineer's deliverable up a layer. Anthropic's S-1, filed October 26, 2026, defines Applied AI teams as the ones that include forward-deployed engineers and says a model alone is not an agent: the harness is the secure environment, credentials, memory, and orchestration loop, and building it was, until recently, work each customer took on itself. The filing now sells that harness as the Agent SDK, the same one that powers Claude Code, and as Claude Managed Agents, which Anthropic operates. So the engineer on site no longer writes the loop, the retries, or the credential handling. What is left is what the filing says is still missing: connectors that do not yet exist, governance frameworks that are not yet written, orchestration logic specific to the customer, and the skills that turn institutional knowledge into agent capabilities. The FDE's output became a skill, a connector, and a row set rather than a repository, which is cheaper to hand back and easier to generalize. The tool also arrives first: the filing says Claude is often already in use by the time its direct teams engage a customer, so week one is an audit of what is running rather than an install. And one thing the tooling did not change. A prompt is code with nothing checking it, as Lesson 08 of Becoming Forward Deployed Engineer put it the same day. Changes got cheap. Verification stayed the FDE's job, and the eval on every change is how it gets done.

The harness used to be most of the job. Anthropic's S-1, filed October 26, 2026, says a model alone is not an agent, defines the harness as the environment, credentials, memory, and orchestration loop around it, and sells that harness as the Agent SDK and as Claude Managed Agents. So the forward deployed engineer's deliverable moved up a layer, to the connectors, skills, evals, and governance the filing admits are still missing per customer. The tool arrives before the engineer does, which makes week one an audit. And nothing in the tooling checks a prompt, which keeps verification in the seat.

Part 13 of The FDE Transition, A Model Alone Is Not an Agent: What the Tooling Changed for the FDE Seat, makes the case from the filing and the week's write-ups by Rebecca Bellan, Matthew Burns, and John Gleeson. Part 4 has the three-layer codebase contract the harness now sits in the middle of, and the handbook chapter on the deployment-to-product loop has the editor's job the harness doesn't take.

Last reviewed 2026-10-29 · 1 min read

What did agent harnesses change for forward deployed engineers?

SOURCES

THE LONG VERSION

RELATED ANSWERS