The Harness Isn’t Dead
The hand-holding is
By Donovan Crewe ·
There is a familiar pattern in engineering. When a new capability arrives, we build systems around it to make it useful. We add instructions, checks, retries and workarounds to compensate for its limitations. As the underlying technology improves, some of that scaffolding becomes unnecessary. What once enabled the capability can eventually become a constraint.
Coding agents are reaching that point.
In February, Amp argued that the coding agent as we know it is becoming obsolete. Newer models can reason, use tools and recover from mistakes with far less prompting and orchestration than earlier generations required. A model with access to something as simple as a shell can accomplish meaningful engineering work without a framework dictating every step. The bottleneck is shifting away from the model and towards the codebase and the environment in which it operates. Amp’s argument
I agree with the direction, but I think the conclusion goes too far. We are not eliminating the harness. We are changing what we need it to do.
Much of the complexity in early coding agents existed because we had to explicitly guide behaviour that models could not reliably manage themselves. As models improve, elaborate prompts, prescribed tool sequences and behavioural workarounds become less necessary. Keeping them in place can introduce complexity and constrain a model that is already capable of finding its own way through a problem.
But reasoning through a task is different from executing it reliably in a real engineering environment. A model can determine how to approach a problem without knowing whether it has permission to change a particular system, which business rules must be preserved or what evidence is required before its work can be accepted. These are engineering decisions, not prompting problems.
A coding assistant typically operates within a feedback loop with a person. An agent can take a goal, inspect a system, make changes, run checks and continue with considerably less intervention. As that independence increases, the environment in which the agent operates becomes a more significant part of the engineering problem.
A July 2026 study of eleven coding-agent systems examines how these environments handle tools, context management, safety controls and orchestration. The variation across the systems reinforces an important point: there is no single harness architecture that makes sense for every task. Study of coding-agent harnesses
For a small, well-defined change, a capable model with access to a repository and a test runner may be enough. A migration across multiple data systems presents a different challenge. The agent must discover what exists, understand dependencies, identify business logic, plan the conversion and establish whether the target behaves as expected. A missing dependency or an undocumented assumption can undermine the result even when the generated code looks correct.
The question is not whether the agent can generate a conversion. It is whether the organisation can verify the result.
In our work on migrations at Lumenalta, we have been exploring how to structure this process so agents can take on more of the execution without losing visibility or control. The workflow we use in Atlus FastPath moves through discovery, planning, conversion, evaluation and reconciliation. Engineers review the proposed plan before execution, evaluations check the work as it progresses, and reconciliation helps identify what has changed and where differences remain.
This structure gives the agent room to work independently while keeping the outcome open to inspection. Rather than relying on a final explanation from the model, engineers can assess progress, identify problems and intervene when necessary.
Across a number of migrations completed so far, we have seen delivery take roughly a third less time, with a fraction of the team that would otherwise have been required. These are early results from a limited set of projects, not a universal benchmark but it clearly proved that combining capable models with a well-defined workflow can change how migration work is delivered drastically.
As models improve, some of the mechanisms supporting this work will become unnecessary. If a model can reliably select the appropriate tools, there may be little value in a specialised mechanism designed to force those interactions. If it can manage context across longer tasks, less manual intervention may be required. Amp has described simplifying parts of its own agent architecture as model capabilities have improved. Amp’s account of rebuilding its agent
The important distinction is between infrastructure that compensates for model limitations and controls that exist to make engineering reliable. Permissions must still reflect the task and the consequences of an action. Long-running execution needs to be observable and recoverable. Evaluations must establish whether requirements have been met, and agents need access to the architectural decisions, ownership and business constraints that shape acceptable outcomes.
Greater autonomy makes these responsibilities more consequential because an agent can take more actions before a person intervenes.
This changes where engineering teams should focus their effort. When an agent fails, the instinct has often been to clarify an instruction, add another rule or introduce another retry. That approach made sense when models struggled to follow complex instructions. It is less useful when the underlying problem is incomplete information, inadequate tests or an environment that does not expose the right tools.
An agent cannot validate a migration if the organisation has no meaningful way to compare the source and target. It cannot make a well-informed architectural decision if relevant knowledge is scattered across disconnected documents and conversations. Better models can make more effective use of the information they receive, but they cannot reliably compensate for information that does not exist or is inaccessible.
For technology leaders, this means an AI strategy cannot stop at choosing a model or adopting a coding tool. The organisation must also consider how agents receive context, where they can operate and how their work is evaluated. This connects directly to Engineering Intelligence: the ability of an organisation to solve problems, learn from outcomes and improve through the combined intelligence of people, AI and organisational knowledge. The opportunity is to make that capability repeatable rather than reconstructing it for every task.
I do not think the future lies in enormous agent frameworks that attempt to anticipate every possible behaviour. Nor will a bare model call be sufficient for every consequential engineering task. We need lighter, more adaptive environments that let models handle more of the reasoning they have learned to perform while retaining the boundaries and verification mechanisms that make their work dependable.
The shift is from engineering every action an agent should take to engineering the conditions in which it can act effectively. Better models give us the opportunity to remove unnecessary complexity. Our responsibility is to ensure that what remains makes their work reliable.
The harness isn't dead. The hand-holding is.