Poesis
Systemic Agentic Framework · Solution Implemented
Agentic Harness — Systemic Agentic Framework.
Give local agentic work explicit steps, artifact checks and traceable decisions.
On supported host adapters, the harness resolves workflow steps and model profiles, supplies instructions, checks mediated writes and records invocations. Checks cover declared schemas and Markdown metadata. Authenticated human approval is planned; atomic writes and cross-host compatibility remain integration boundaries.
Values
Value delivered by the product
- Value Partial
Inspect control decisions.
Review persisted step and write-authorization decisions within mediated execution, alongside separate checks for output quality and human approval.
Supported by Affordances(1)
Supported by Capabilities(1)
Supported by Features(2)
- Value Partial
Catch invalid declared artifacts earlier.
Check matched schemas before mediated writes so declared structural errors surface early. Coverage follows the supported parser, contract and transaction boundaries.
Supported by Affordances(1)
Supported by Capabilities(1)
Supported by Features(1)
- Value Implemented
Reuse methods on supported hosts.
Carry plain-file methods between supported host/model adapters, with explicit compatibility and mediation boundaries.
Supported by Features(1)
Realized through Use cases(1)
Features
Implemented and planned features
- Feature Implemented
Workflow & step resolution
Resolve eligible workflow steps from configuration and journal outcomes on supported adapters, with dispatch handled by the host.
Addresses Pain points(1)
Supports Values(1)
Serves Use cases(1)
Used by Actor types(2)
- Feature Implemented
Skill and instruction prompt injection
Supply resolved instructions and skill-loading directives through supported adapters. Qualify delivery of pinned guidance and declared checks on target hosts; receiving instructions does not establish model adherence.
Addresses Pain points(1)
Supports Values(2)
Serves Use cases(1)
Used by Actor types(2)
- Feature Implemented
LLM resolution
Select model profiles by task capability and configured cost ranking, with slug tie-breaks and adapter mappings. Qualify routes using retained workload-specific quality, cost and disclosure evidence supplied by evaluators; ranking alone does not establish suitability.
Addresses Pain points(1)
Supports Values(4)
Serves Use cases(1)
Used by Actor types(2)
- Feature Implemented
Artifact validation
Check mediated write sets against matched schemas before writing. Scope: structured data and Markdown metadata; prose quality, contract resolution and atomicity have separate boundaries.
Addresses Pain points(1)
Supports Values(1)
Serves Use cases(1)
Used by Actor types(2)
- Feature Implemented
Logging
Trace applicable harness invocations in correlated per-session JSONL journals. Coverage is limited to captured calls; synchronization, integrity and recovery are separate controls.
Addresses Pain points(1)
Supports Values(1)
Serves Use cases(1)
Used by Actor types(2)
Roadmap
Where Agentic Harness is going.
Recorded version: v1.0. Historical release labels do not establish the current feature scope above.
- v1.0 Implemented Recorded release: MVP/GA — resolution, injection, LLM routing, validation, authorization, logging
Implemented and Planned qualify the features in each milestone. Recorded releases and planned milestones preserve history, not delivery dates or hosted availability. See the Milestones & Versions register.
Qualities
What Agentic Harness claims about how it behaves.
- Qualified model routes Partial Ranking alone does not establish suitability
Model profiles are selected by task capability and configured cost ranking. Routes are qualified using retained workload-specific quality, cost and disclosure evidence.
- Disclosure limits Planned What a route may be told is bounded
Instruction delivery and disclosure limits are evaluated per workload. A host change is a change of behavior, not a configuration detail.
Usage
Use cases involving Agentic Harness
- Use case Implemented
Approve an agent resource and action mandate
Review resource/action grants against a bounded assignment and its expiry, then check those grants and the target baseline before mediated work. Mandate review is planned; grant checking alone is not authenticated human approval.
Supported by Features(1)
Addresses Pain points(1)
Realizes Values(6)
Performed by Actor types(3)
- Use case Implemented
Qualify model routing for a workload
Compare candidate model routes against workload-specific quality, cost and disclosure criteria using retained evaluation evidence.
Supported by Features(1)
Addresses Pain points(1)
Realizes Values(4)
Performed by Actor types(2)
- Use case Implemented
Qualify instruction delivery on another host
Verify that a target host receives the pinned instructions and declared checks, distinguishing delivery from model adherence before migration.
Supported by Features(1)
Addresses Pain points(1)
Realizes Values(2)
Performed by Actor types(2)
- Use case Implemented
Resolve the next eligible workflow step
Resolve eligible steps from workflow configuration and recorded outcomes on a supported adapter; leave dispatch to the host rather than treating resolution as execution.
Supported by Features(1)
Addresses Pain points(1)
Realizes Values(1)
Performed by Actor types(2)
- Use case Implemented
Validate a mediated artifact write
Check a proposed write set against a matched schema and inspect structured-data or Markdown-metadata violations before writing. Passing validation does not establish prose quality or acceptance.
Supported by Features(1)
Addresses Pain points(1)
Realizes Values(1)
Performed by Actor types(2)
- Use case Implemented
Inspect a captured harness invocation sequence
Read correlated session-journal records to reconstruct captured harness invocations and their outcomes, distinguishing absent records from evidence that no action occurred.
Supported by Features(1)
Addresses Pain points(1)
Realizes Values(1)
Performed by Actor types(2)