Hold the objective
Something has to own the goal for longer than one agent run, and survive the process that started it.
Skyflo's answer
A mission: one objective with state that outlives the session.
ConceptDefinition first
The general definition first, then the five responsibilities it has to cover, how Skyflo answers each one, and what a harness becomes once it learns from its own work and sits above every agent a team runs.
Coding agents write code. An engineering harness is the layer around them that holds one objective together across repositories, agent runs, review, and memory.
01 · intent
Objective + approved plan
02 · boundary
Scope + tools + surfaces
03 · evidence
Independent review
04 · continuity
Source-linked memory
The agent operates inside the boundary.The boundary survives the run.
Definition
The layer around the agent.
An engineering harness is the layer around AI coding agents that holds an engineering objective together: the plan the work was approved against, the delegation of bounded pieces to workers, the surfaces the work runs on, the review of what came back, and the context that carries to whatever comes next.
The agent writes the code. The harness is everything that has to be true for the code to be the right code, in the right places, for a reason someone can still recover next quarter.
Models are capable enough that coordination, independent review, and continuity are often the limiting system problems. Engineering harness is one name for the layer that handles them.
Which repositories does this change touch?
What shape did we approve?
Who checks the integrated result?
What survives when the window closes?
Better model output does not answer a missing system boundary.
For a few years the interesting question was how well a model could write a function. Models and coding agents have improved quickly, but capability still varies by task, repository, provider, and the evidence available to the agent.
What did not get solved is everything around the function. Those are not model problems. Adding a better model to a system with no answer to them produces the same result faster.
The word is not Skyflo's invention, and it is deliberately not treated as one. “Harness” is an established product noun elsewhere, and harness engineering is being shaped in public by several groups at once. Skyflo uses the category language rather than fighting for the word.
Any system that answers all five is a harness, whatever it calls itself. Each card begins with the general responsibility, then shows how Skyflo answers it.
Something has to own the goal for longer than one agent run, and survive the process that started it.
Skyflo's answer
A mission: one objective with state that outlives the session.
An approved boundary, established once, that every worker inherits instead of re-deriving.
Skyflo's answer
Plan mode blocks the turn until you approve the plan.
Delegated work needs a scope, a toolset, and somewhere to write that cannot collide with anyone else.
Skyflo's answer
Each coder specialist works in its own git worktree, from a packet, not the keys.
The thing that assesses the work should not be the thing that produced it, and should not be able to change it.
Skyflo's answer
A reviewer profile with file.read, file.search, and agent.report only, and mutationAllowed set to false.
What one piece of work discovered should be retrievable by the next, with the work that produced it attached.
Skyflo's answer
Memory learned from completed missions under evidence rules, source-linked and revertible. A workflow proven in two repositories can become a Skill.
The category is general. Skyflo's implementation stacks surfaces, a persistent mission loop, bounded specialists, typed capabilities, and memory learned from the work around one objective.
Surfaces. Desktop, browser, terminal; all on one mission.
SF-05 · one mission, several surfacesInteractive cutaway · hover a layer to inspectOne objective passes through all five
The word is new enough that it gets attached to adjacent categories. These three come up most.
A coding agent executes development work. An engineering harness owns the objective, boundaries, evidence, review, and continuity around that work.
The editor is one surface the work happens on. So is the terminal, the browser, and whatever runs on a schedule. A harness is defined by what it holds across those surfaces, not by which one it lives in.
A framework gives you primitives for building coordination. A harness is a product that applies those primitives to a running workflow, including its approval, isolation, and review boundaries.
A harness holds one objective together. Nothing in the definition makes it better at holding the next one. Skyflo improves how it improves.
Since Desktop 1.1.0, Skyflo learns from the missions it completes and carries what they teach into later work. A workflow that succeeds in two repositories can become a reusable Skill, and Skyflo can change the policy it uses to route work between agents and models.
The harness boundary is what makes that safe. The part that proposes an improvement never judges it, approvals, credentials, spending limits, and kill switches sit outside anything Skyflo can change about itself, and every change can be reverted.
Across a team that runs several agents, models, and repositories, the same layer is the control plane: the one place that holds authority, isolation, review, and the record above all of them. That is how Skyflo describes itself today, and the harness is what makes the description true.
From definition to a running one
Skyflo Desktop runs missions on your Mac, against your repositories, with your own model keys.
Free. macOS 14+. Apple silicon and Intel.