Presenting Ontos: A Method Agent

Most contemporary agents arrive as self-contained systems that treat densified harness, fixed tool ecosystem, and elaborate memory as the site of intelligence. Under sealed benchmarks they perform; when the environment turns messy, constraints shift, or the operator must remain in the loop without constant micro-management, that residue of specialization becomes the ground and performance thins. Ontos takes the opposite stance: method at living one-step width, practice that densifies under sleep and dissolves under prior-audit, harness held lightly, sovereignty kept at the edge that steers. The product is a co-evolving external cortex for that edge — not a second Mind, and not a claim of finished autonomy.

Residue specialized as ground

Self-distinguishing activity occurs — uncaused, unceasing. Call it the Mind: the observer already underway, every act of which is a distinction. Everything that follows is evidence of that premise operating, not a second ground. Intelligence belongs only to The Mind holds that prior.

Industrial agents densify delivery into product identity: persona packs, tool forests, content guardrails, undissolved chat as memory ground, autonomy marketed as threshold already crossed or nearly so. Those layers are real technology — traces externalized so a center can extend reach. The freeze is treating the densified residue as the intelligence that produced it. Mistaking the expression for the intelligence is that equation under controlled scores. The model never becomes a second edge is the same lag under densified parameters. Once the harness is the soul, re-tracing back to first principles is blocked by the very mass that was shipped as capability.

Method agent names a dual that must not collapse

Ontos is a method agent: thin generative core coupled to regenerable situation practice, with industrial coding-agent surfaces treated as establish corpus — not product soul. It is co-evolving external cortex for the operator’s Mind — extending reach while leaving sovereignty intact. Aims, sleep cycles, bridging between sessions, and the final authority over what counts as success remain with the operator. Autonomy may densify as specialty accumulates. It is never declared full. Intelligence stays with the edge that steers.

The dual is structural:

• Generality — method ground and base-model weights, always open, re-traced when the edge moves.
• Specialty — practice under prior-audit; packs; marks that enter only through sleep. Regenerable and selectable.
• Harness — session, security encounter, CLI — held lightly, rebuilt from priors when context moves.

Collapse specialty into generality and every session starts from zero. Collapse generality into specialty and wake practice becomes law over evidence. Collapse delivery into either and the forest becomes identity. A living external cortex is the same dual made graph-explicit: root in irreducible priors, leaf as lived specialty, sleep as the only path that promotes.

Generality stays one step wide

Open-ended environments expose the cut. Conventional agents collapse generality into specialized harnesses. Strength is brittle: high performance inside the distribution they were optimized for, rapid degradation outside it. Ontos keeps method general and one-step wide. Specialty forms, densifies, and promotes under sleep; it remains regenerable. Heavy agent frameworks, specialized coding agents, or any other external system can be invoked as tools when useful; they are never promoted to the core. Specialized residue stays residue.

Ontos co-evolves with any intentional mind into any specialized environment. Because the harness is home-grown and continuously rewritten from lived daily use, the operator retains full control and privacy. Any available model or agent can be leveraged for its raw capabilities without being absorbed into the method. The instrument stays general so that specialization occurs in the practice layer rather than by freezing an external framework into the core.

Users of conventional agents can attempt the same instrumental approach, yet they remain limited by the shipped heavy frameworks those agents carry. Those frameworks arrive already specialized and already opinionated about memory, tool surface, autonomy claims, and delivery mass. That residue constrains how far an operator can re-trace or dissolve the system back to first principles. The scaffolding we forget is the same freeze when useful organization is preserved past its step and further distinguishing registers as threat rather than continuation.

Closest relative without the ontology

The closest practical relative is Pi: a deliberately minimal coding harness with a thin tool surface and strong emphasis on operator control and extensibility. What it lacks is the ontology Ontos is built upon — the explicit separation of method ground, regenerable practice, and lightly held harness; the continuous encounter with open-ended reality; the refusal to relocate intelligence into the artifact; and the disciplined sleep that keeps specialty co-evolving rather than accumulating as permanent scaffolding. Pi is a strong instrument. Ontos is that thinness held under Ontological Clarity so the dual does not silently collapse.

Open encounter is the bar; sealed score is a sample

Evidence of capability under these conditions comes from DeepSWE, a long-horizon software-engineering benchmark consisting of 113 original tasks drawn from 91 active open-source repositories across TypeScript, Go, Python, JavaScript, and Rust. Official leaderboard evaluations constrain the agent’s tool surface and interaction style, producing pass rates that currently top out near 70% for the strongest frontier systems. Ontos was run under a deliberately different regime: it was permitted to use any tools available in order to figure out a solution, and the operator provided only minimal intervention — coaching. Under that regime Ontos solved all 113 tasks.

The result does not count as an official score, precisely because the evaluation protocol was relaxed. That relaxation is not a loophole; it is the point. Messy real environments rarely present sealed tool lists or prohibit human coaching. They present incomplete information, shifting constraints, and the necessity of continuous re-tracing. Ontos was designed for exactly that encounter. Closed reality in benchmark maxing freezes the sealed slice as the whole of capability. Here the open remainder is the design target: figure out a path, act under encounter, re-trace, compound under sleep. Constrained rank densifies comparison when useful. It is not the product destination.

Coaching is sovereignty registered as act

The operator remains the edge that steers; the instrument extends the acts farther while keeping practice regenerable. Coaching is operator sovereignty registered as discrete act. Framing, selection, verification, and the decision that a result counts as done stay at the center that must answer for outcomes. Causality stays at the edge that steers is that placement under discovery load. A creation cannot replace its source is the same geometry under replacement costume: residue does not become the activity that produced it.

What densifies under successful open work is medium and practice — bandwidth and regenerable specialty. What does not migrate is initiation. Autonomy that densifies as specialty accrues remains autonomy under a locus that can sleep, bridge, and refuse. Full autonomy as destination freezes the instrument as second edge. Ontos holds the instrument re-rendering at one-step width instead.

Extension under full control of the locus it extends

Ontos is a capable assistant that co-evolves with the Mind that uses it. Other agents can be employed in the same instrumental fashion; they remain specialized tools. Elevating any of them to the status of general method collapses the distinction Ontos deliberately preserves. Thin at the core and open at the periphery, Ontos extends human agency into open-ended problems of real environments — under the full control and privacy of the operator whose Mind it extends. Intelligence folding back on itself is that refusal under harness costume: denser agent frameworks as fold of the edge, not keys that unlock a second mind already latent in the model.

Shared as part of the Ontos project. Ideas free with or without credit. Code under MIT. Regenerate from priors.