Skip to content

reference-agent-patterns.md

Latest commit

 

History

History
68 lines (54 loc) · 3.23 KB

reference-agent-patterns.md

File metadata and controls

68 lines (54 loc) · 3.23 KB
Raw

Reference-Derived Agent Architecture

Palamedes's agent runtime should be assembled from locally available, inspectable reference implementations rather than from a single generic agent framework.

Adopted patterns

Concern Primary reference Pattern to adopt Palamedes boundary
Durable direction memory Palamedes Plan, evidence, hypotheses, revisions, restore Owned by the kernel
Wake and work ownership Paperclip Heartbeat/wake context, explicit ownership, budgets, idempotent approval Owned by the host
Tool safety AgentScope Per-tool permission decisions, sandbox/workspace isolation, event stream Owned by the host
Within-cycle memory Dexter Scratchpad, bounded tool loop, context compaction, separate final synthesis Owned by the cycle runner
Specialist behavior gstack Role-specific skills, real-environment QA, context save/restore, eval persistence Skills and evaluation layer
Extensibility OpenClaw Lean core, plugins for optional capability, explicit channels and safe defaults Plugins around the host
Shared team cognition Palamedes team ledger Provenance, observation bias, competing hypotheses, single-owner missions, outcome attribution Epistemic state only; host still schedules

Canonical source locations are recorded in /Users/ze/work/ref/manifests/status-current.json; automated keyword-level coverage is stored in /Users/ze/work/ref/catalog/agent-patterns.json.

Minimum vertical loop

wake
  -> read Palamedes snapshot
  -> AI strategist report
  -> persist grounded reference insights
  -> validate next_actions against capabilities
  -> execute bounded host actions
  -> stop on failure, limit, or human review
  -> return the post-cycle snapshot

The implementation is runtime/agent_cycle.py in the agent scaffold. It is deliberately bounded and does not add a queue, scheduler, shell executor, or manager-of-managers hierarchy to the Palamedes kernel.

For 1:N teams, the companion team cognition contract supplies shared world-state versioning and plural hypothesis memory without moving orchestration into the kernel.

Rejected defaults

  • Unbounded autonomous loops
  • Silent permission bypass outside a sandbox
  • Multiple manager layers before a single-agent loop is proven
  • Persisting model prose without source IDs, URLs, or evidence quotes
  • Treating structural schema validity as evidence of decision quality
  • Copying optional integrations into the core instead of exposing plugin hooks

Evaluation contract

The initial comparison dataset is scaffolds/palamedes_agents/evals/agent-cycle-cases.json. A valid release claim requires three real project cases, at least two blinded preferences over a generic baseline, and at least one attributable stop, pivot, or positioning decision. Unit tests and schema checks are necessary but do not satisfy this product-quality gate.

Ouroboros boundary comparison

Ouroboros is tracked as a contrasting execution-oriented reference. The adopt/adapt/reject review is recorded in reference-ouroboros.md. Palamedes borrows its execution-versus-evaluation distinction and bounded contract patterns, but not its Agent OS execution loop.