Skip to main content

The beyond10x agent loop · pre-v1

One loop.
Explicit effects.

Give a model a task, decide what it can touch, and keep a record of every turn. Harness talks to model APIs directly, asks before a tool call above the run's risk ceiling, and holds the run to the bounds you set.

Rust · OpenAI Responses and Anthropic Messages · Linux · LicenseRef-B10x-Proprietary

shipped, runs today
35
planned
2

01 / Start small, add authority deliberately

A capability is a boundary, not a promise in a prompt.

Begin with a read-only workspace. File changes and process execution appear only when a confinement boundary can provide them, and a missing capability is reported as withheld.

  1. Read the workspace

    Shipped

    file_read, dir_list, search and find, bounded to the workspace root. The default needs no substrate and publishes no write tool; b10x-harness tools prints the catalogue without contacting a model.

  2. Admit a change

    Shipped

    A confined workspace adds file_write and file_edit. Each call still meets the run's approval ceiling and its ordered write scope.

  3. Allow execution

    Shipped

    run starts an argv, never a shell string, and only for a program declared with --allow-program, inside confined execution. A host that cannot confine it withholds the tool.

  4. Budgets that bind

    Shipped

    Turn, token, duration and context-window ceilings, and a spend ceiling when a dated rate card prices the model. Unreported usage stays unknown and is never counted as zero.

  5. A readable record

    Shipped

    Sessions are written outside the workspace and resume locally; --json streams vendor-neutral JSON Lines events for another program.

  6. Driven by a real external client

    Planned

    Bridge mode serves the loop over JSON-RPC on stdio and is contract-tested, but no real external client has driven it yet.

02 / One turn

The model proposes. The loop decides.

Every turn is assembled from the local conversation and sent as a complete request. A refusal comes back to the model as a failed tool outcome, so it learns the effect did not happen.

  1. HarnessAssemble

    Instruction, conversation, published tools and the budget left.

  2. Model APIPropose

    A stateless request over OpenAI Responses or Anthropic Messages.

  3. HarnessGate

    Catalogue, risk ceiling, approver and write scope, per call.

  4. WorkspaceExecute

    Admitted tools run; each result is bounded and returns to the next turn.

  5. HarnessRecord

    The answer, the session and a typed stop when a bound ends the run.

03 / Status

Pre-v1. Here is exactly what exists.

Shipped means code on main today, held by the named test. Planned means not built yet. Provider fixtures come from deterministic emulators; dated live observations are listed under Limitations. Generated by harness-docs.

Shipped · runs today

  • ShippedOpenAI ResponsesThe binary reads a real file through a real tool call over the Responses wire.
  • ShippedAnthropic MessagesThe binary drives the Messages wire and calls a real tool through it.
  • ShippedProvider presetsA short model name resolves to the exact identifier this build pins.
  • ShippedCredentials the operator namesNaming your own credential turns the preset's token renewal off, so no file is rewritten that the operator did not offer.
  • ShippedSecrets store only on requestA run or server without the secrets flag never dials the session bus.
  • ShippedTurn ceilingA turn ceiling stops the loop, and the stop names the ceiling.
  • ShippedSpend ceilingA spend ceiling on a run that cannot price its model is refused before the first request.
  • ShippedUnpriced is not freeA model the rate card does not price has no cost rather than a zero one.
  • ShippedUnreported usageUsage a provider did not report stays unknown rather than becoming zero.
  • ShippedCompactionA declared context window compacts on the token count the provider reported.
  • ShippedCancellationCancellation between tool calls stops the run before the next effect.
  • ShippedStructured outputA run asked for a schema that answers in prose stops without an answer.
  • ShippedDeny by defaultA run entry built from the process adapter is refused under the default approver.
  • ShippedNo silent approvalA write is refused when the run has nobody it may ask.
  • ShippedApproval ceilingRaising the unattended ceiling lets a write through, on either wire.
  • ShippedWrite scopeA path the write scope denies is refused for both writers, and the refusal says the scope is what is wrong.
  • ShippedWithheld tools are reportedA declared program with no confinement is reported as withheld rather than dropped.
  • ShippedVerb surfaceThe verbs surface publishes three tools over the same catalogue.
  • ShippedOutbound MCPAn MCP server named in a reviewed profile runs end to end over stdio.
  • ShippedToolset descriptionThe tools command describes the toolset without an endpoint.
  • ShippedProfilesEvery profile that contributed to a run is in its record with its own digest.
  • ShippedDefault write scopeThe default write scope denies the repository's own Git directory.
  • ShippedDeclared context filesA declared context file that is absent refuses the run before any session is filed.
  • ShippedResumeA run files its conversation, and a later run resumes it.
  • ShippedSessions stay outside the workspaceA session directory inside the workspace is refused before it is created.
  • ShippedJSON eventsJSON mode emits one event per line.
  • ShippedDelegationA delegate runs a whole second loop, and the parent reads only its report.
  • ShippedConcurrent delegatesTwo delegates of one turn run at the same time.
  • ShippedNamed agentsA named agent cannot widen what the parent run was admitted.
  • ShippedSkillsSkill descriptions are in the instruction, and their bodies are not until the model asks.
  • ShippedRead-only memoriesrecall answers a memory's body, and the record names every memory whatever its status.
  • ShippedStop hooksA stop hook blocks one ending, and the run turns again before it finishes.
  • ShippedHook pointsA hook file naming a point that does not exist is refused by name.
  • ShippedTool calls round tripA tool call round trips back to the client that registered the tool.
  • ShippedPinned method inventoryThe client methods served and refused match the pinned profile manifest.

Planned · roadmap only

  • PlannedDriven by a real external clientA real external client drives bridge mode end to end, beyond the contract tests.
  • PlannedThe loop in the executable specificationThe agent loop's turn, tool and budget behaviour is modelled in the executable specification and run by its conformance suite.

As of 2026-10-22 · harness-docs, from its capability list; every shipped item names the test that holds it