Skip to content

Project notice

1.35.0 — Computer Use and retained strategy development

Version 1.35.0 makes valuable capability development through Computer Use the primary path in the English/French SUCCESSOR Ω flagship. Start with a software, research or mathematical opportunity, develop competing strategies, challenge the original formulation and retain useful work from the configured runner. The catalog example remains available without a model, account, wallet or installation. The matching published manifest, exact-source acceptance and served-site checks identify delivery; preparing a source version is not evidence of publication.

Included

  • English/French mission briefs cover valuable digital-work opportunities, open mathematical questions, public software bottlenecks and decisive research experiments. Each has its own acceptance conditions, strategy hypotheses and matching work surface. The homepage features this journey; catalog reconciliation remains an optional local worked example. Missing evidence produces an investigation plan, not an invented result.
  • A guided Computer Use mission with an explicit opportunity, beneficiary/value hypothesis, acceptance criteria, competing strategies, a proposed combination, an alternative formulation, input rights and finite execution limits.
  • A real public-software starting mission: inspect the released Computer Use runtime and deliver a source-linked patch proposal or a reason to retain the design, with a reproducible before-and-after validation plan. Research and mathematics lanes request concrete mechanisms, falsifiable experiments or checkable candidate arguments. They do not substitute a fictional product comparison for the user's intended work.
  • Level 3 — meta-strategic intelligence: discover strategy candidates, select and test a baseline or combination, retain observations and propose an improved method. The executable goal asks the model to distinguish methods actually tested from untested ideas and to produce a visible work product.
  • Level 4 — transformational intelligence: challenge an unnecessary restriction and test a different formulation while preserving the original outcome and acceptance criteria. A reframing remains a hypothesis until its actual result supports it; it cannot silently expand permissions or relax requirements.
  • Project-bound mission packs accepted by the existing successor-computer prepare/run interface. Exact project bytes, source identity, strategy and attempt UUID are bound into the existing task/provider/image/worker permit. Browser preparation makes no model call and carries no credential.
  • Portable native returns retain observed provider/model identity, bounded actions, usage observations, elapsed time, the existing signed receipt and final visible text when available. Interrupted sessions preserve reservations and uncertain effects; the native runtime never automatically retries them.
  • One project history for prepared Computer Use attempts, imported returns, criterion-by-criterion owner reviews, review effort, catalog cycles and retained learning. Compare attempts only within the same outcome, work surface, source, rights and limits. Keep unsuccessful attempts as useful history.
  • A complete project archive and portable offline workspace preserve the combined history. Existing version-one catalog/learning imports remain supported; archives with Computer Use history use the additive learning-archive version two. Oversized archives are refused rather than silently pruned.
  • Fail-closed provider completion: empty, reasoning-only, refusal-only and commentary-only Responses output cannot complete a session. A nonblank final message remains an unverified completion claim.
  • The native execution goal retains the complete strategy slate while permitting only the selected approach. Current Responses pointer actions with empty modifier lists are accepted without broadening supported controls.
  • Public homepage and flagship entry points introduce the screen-based mission while preserving the free local catalog path, existing laboratories, research, diagrams and separate USDC/$AGIALPHA routes.

Execution and qualification limits

The native actuator reads approved public sources and creates a visible work product in disposable Chromium inside an isolated Linux Docker container. Its current Responses computer interface and alternate compatible vision/JSON supplier require explicit native configuration. The operator selects an available model and immutable image; a provider or model name does not independently establish task fitness. The static website does not itself operate a remote computer. The native guide identifies current task-fit model options with primary references. This runtime does not connect arbitrary code execution, performance measurement or a formal prover. A patch proposal is not a tested patch; a candidate argument is not a verified theorem.

This release scope excludes full desktop control, private browser profiles, customer credentials, private data, account login, arbitrary model-supplied code, payments and message sending. Exact approved public HTTPS reads remain bounded by the existing broker. Existing stop, reservation, reconciliation, cleanup and quarantine boundaries remain in force. Token observations are not an invoice; unknown billing stays unknown.

Local control tests use explicitly labeled supplier/actuator doubles. Browser acceptance and cross-language handoff checks validate software behavior and consistent bytes. They do not establish a completed live-model task, production Docker isolation or independent acceptance of a paid outcome. Native Docker and live-provider qualification for this new flagship path remain separate operational gaps: run authorized exact-source container/provider trials, retain screenshots and request/usage evidence, record costs and review the actual deliverable before claiming those outcomes. Full required hosted CI and public-site acceptance must independently pass for publication; their status is established by the matching workflow and manifest, not this document.

The browser checks historical return consistency; it does not install a trusted signer from imported data. Verify native signatures against a separately trusted source key. An owner's useful-result review remains that owner's judgment, not independent certification, Specialist admission, inherited authority or settlement. The broad screen-based-work ambition does not establish universal capability or a measured economic advantage. The proposed $40 trillion/year market is the owner's estimate, not an independently established result.

Continue from 1.34.0

Version 1.34.0, its tag, assets and deployment evidence remain unchanged. Retain earlier downloads and project exports, obtain the new release's matching assets in a new destination and inspect the manifest/checksums. The published catalog workflow and its native workbench returns remain distinct from configured model-driven Computer Use. An older reader cannot interpret the additive learning-archive version two; use the matching new workspace to restore combined Computer Use history, while preserving the original exports.

No contract deployment, wallet migration, funding, activation or provider account is commissioned by this software release. Current Mainnet records and historical wallet profiles retain their pinned identities. AGIJobManager commissions USDC work; the Alpha / Jobs v0 contract route uses $AGIALPHA. Each retains its own configuration, budget, review, authorization and settlement requirements.

First-use guide · Flagship guide · Native Computer Use guide · Release readiness