Computer-use agents from Anthropic and OpenAI point a frontier model at the screen and let it reason its way through a task. That flexibility is real and improving fast. OpenAdapt compiles a demonstration once and replays it deterministically, reserving models for compilation and repair.
Every claim below names its public source. Where a vendor publishes no figure, we say so instead of estimating one.
OpenAdapt
Healthy runs replay the same compiled steps deterministically with zero model calls; drift outside what was demonstrated halts for a governed decision.
computer-use agents
Every run re-reasons from screenshots: OpenAI's loop sends a screenshot back after each action batch so the model can plan the next step, which adapts to novelty but makes each run non-deterministic by construction.
OpenAdapt
$0 model cost per healthy run on the MIT runtime - published trials across MockMed triage (18 trials), Frappe Lending (6/6), and openIMIS eligibility (6 runs) all recorded zero model calls.
computer-use agents
Metered per model turn: screenshots are billed as image inputs every step. Claude's computer-use toolset definition alone adds about 4,500 input tokens to a request before any screenshot; OpenAI recommends original-detail screenshots each turn for click accuracy.
Source: Claude computer use tool: pricing · OpenAI computer use guide: screenshot detail guidance · OpenAdapt published trials with 0 model calls
OpenAdapt
Every consequential run ends VERIFIED or HALTED based on an independent check of the system of record, with silent incorrect success counted as a tracked failure metric in published evidence.
computer-use agents
The acting model reports task completion from what it sees; providers direct developers to keep a human in the loop for high-impact actions rather than providing an out-of-band business-effect oracle.
Source: OpenAI computer use guide: human-in-the-loop safeguards
OpenAdapt
Halting is a designed outcome: ambiguity, wrong identity, or a refuted effect stops the run with preserved evidence instead of proceeding or retrying blindly.
computer-use agents
The model keeps attempting until it believes the task is done or its budget is exhausted; stopping criteria are prompt- and harness-level, not contract-level.
OpenAdapt
Runs execute locally inside your boundary; egress is an explicit observed flag (false in the published demo pack), and Claude computer use is documented as ZDR eligible when your application controls storage.
computer-use agents
Screen content leaves the machine to the model provider on every step: OpenAI's reference loop ships screenshots to the API each action batch, and Anthropic bills those screenshots as image input under standard API data handling.
Source: Claude computer use tool: data retention and ZDR · mockmed-triage-v3 pack: off-box egress observed false
OpenAdapt
Browser in production today plus native desktop and external-lane remote execution into managed RDP/Citrix estates, with published acceptance evidence per surface.
computer-use agents
Any UI a screenshot can describe, including desktops inside VMs or containers the harness controls; consumer agents such as Operator extend the same approach to end users.
These are the differences that hold up under scrutiny. Recording demonstrations, visual targeting, virtual desktop awareness, and selector repair are broadly available across modern tools and are not claimed here as unique.
A run is judged by an out-of-band check of the system of record, such as a read-only API call, a SQL query, or re-reading the persisted record, not by the acting session declaring itself successful.
Every consequential run ends verified or halted with a preserved run report. There is no silent third state where the workflow looked finished but the record never changed.
A compiled workflow replays deterministically with zero model calls on healthy runs. Model spend is reserved for compilation and reviewable repair.
For managed Citrix, RDP, and VDI estates, OpenAdapt can drive the local client window from outside the session, so nothing is installed inside the remote environment. The lane is qualified today against a deterministic stand-in and a real FreeRDP round trip; a real ICA/HDX environment is qualified per customer before consequential use.
Recordings, screenshots, and compiled bundles can stay inside your boundary. Local, self-hosted, and customer-controlled deployments are first-class, not an enterprise afterthought.
The compiler and governed runtime are MIT-licensed and inspectable. You can audit exactly what runs beside your systems of record.
Each execution surface ships with bounded, published acceptance evidence, counted effects, refusals, and halts, instead of an unbounded compatibility claim.
Per-surface acceptance results are published in the qualification evidence.
They are complementary. Agent providers themselves recommend isolated environments and human oversight for high-impact actions; OpenAdapt uses models too - at compile and repair time - while healthy runs replay deterministically with zero model calls.
When correctness matters, replay a verified program: OpenAdapt's published evidence counts verified outcomes, expected halts, and silent incorrect successes per workflow - for example 18 MockMed triage trials with 0 model calls and 0 silent incorrect successes - instead of re-reasoning the task on every run.
Source: mockmed-triage-v3 evidence pack
Choose a computer-use agent for exploratory, novel, or constantly changing work, for research and triage, or for tasks you will run a handful of times and never again.
Choose OpenAdapt when the same consequential workflow repeats: healthy runs are deterministic with zero model calls and zero per-run token cost, every run ends verified or halted against an independent check of the system of record, and the runtime is MIT-licensed and can execute entirely inside your boundary.
These approaches are complementary rather than rivals: agent providers themselves recommend human oversight for consequential actions, and OpenAdapt uses models too, at compile and repair time rather than on every healthy run. The question is whether each run should re-reason the task or replay a verified program.
Bring one repeated, consequential workflow and measure authoring time, run time, intervention rate, and incorrect-success rate against your current approach. Or watch the governed-execution demo first: every run shown ends verified or halted.