Skip to main content
EVOKORE// SKILLS / functionality-and-workflows
Sign in
Workspace
Agent-33markdownsha E11944AD0F77

functionality-and-workflows

documentation

Functionality and Workflows

This is the current behavior map for AGENT-33 on main as of February 15, 2026.

1. Runtime Architecture

Primary entry point:

  • engine/src/agent33/main.py

Startup wiring initializes:

  • PostgreSQL-backed LongTermMemory
  • Redis client
  • NATS bus
  • Agent registry from engine/agent-definitions
  • Model router + embedding provider/cache
  • BM25 index + hybrid searcher + RAG pipeline
  • Progressive recall
  • Skill registry + skill injector
  • Code executor bridge for workflow execute-code action

Routers mounted:

  • Health, chat, agents, workflows, auth, webhooks, dashboard, memory, reviews, traces, evaluations, autonomy, releases, improvements, training

2. Capability Inventory

DomainStatusNotes
Chat completionsOperational/v1/chat/completions proxies to Ollama with injection scan
Agent registry + invokeOperationalRegistry discovery from JSON definitions; invoke supports skills + progressive recall injection
Workflow engineOperationalSequential/parallel/dependency-aware modes; 8 step actions (invoke-agent, run-command, validate, transform, conditional, parallel-group, wait, execute-code)
Memory + RAGOperationalVector + hybrid retrieval; progressive recall levels
Tool frameworkPartialTool registry/governance exists, but full runtime bootstrap is not fully wired in main.py
Code execution layerPartialCodeExecutor wired, but starts with tool_registry=None; adapters must be registered explicitly
Review automationOperationalIn-memory two-layer signoff lifecycle
Trace + failure pipelineOperationalIn-memory trace collector and failure records
Evaluation + regression gatesOperationalIn-memory evaluation runs, baselines, regression recorder
Autonomy budgetsOperationalIn-memory budgets, preflight, runtime enforcement, escalations
Release automationOperationalIn-memory release lifecycle, sync rules/executions, rollback records
Improvement operationsOperationalIn-memory intake, lessons, checklists, metrics, roadmap refresh
Messaging webhooksPartialRoutes exist; adapters must be registered explicitly
Training APIsPartialRoutes exist; default startup only initializes training_store, not full runner/optimizer wiring

3. Persistence Boundaries

Persistent (database/backing service):

  • memory_records via LongTermMemory (PostgreSQL + pgvector)
  • Workflow registry and execution history (api/routes/workflows.py) via OrchestrationStateStore when orchestration_state_store_path is configured
  • Workflow run archives, replay events, and extracted artifacts (workflows/run_archive.py) via the file-backed archive under workflow_run_archive_dir
  • Review records (review/service.py) via OrchestrationStateStore when orchestration_state_store_path is configured
  • Trace/failure records (observability/trace_collector.py) via OrchestrationStateStore
  • Autonomy budgets/enforcers/escalations (autonomy/service.py) via OrchestrationStateStore
  • Release records/sync/rollback state (release/service.py) via OrchestrationStateStore
  • Improvement intakes/lessons/metrics/checklists/refreshes (improvement/service.py) via pluggable learning signal stores
  • Approval tokens, tool approvals, process state, and mutation audit records via their configured state stores
  • Training store tables (when training_enabled and initialized)
  • Redis ephemeral cache/state
  • NATS event transport

In-memory only (lost on restart):

  • Evaluation runs/baselines/regressions (evaluation/service.py + recorder)
  • Auth users and API keys (api/routes/auth.py, security/auth.py)
  • Live workflow WebSocket snapshots/queues (workflows/ws_manager.py) before they are flushed into the run archive

4. Workflow Lifecycles

4.1 Review Lifecycle

States:

  • draft -> ready -> l1-review -> l1-approved -> (optional l2-review -> l2-approved) -> approved -> merged

Main APIs:

  • /v1/reviews/{id}/assess
  • /v1/reviews/{id}/assign-l1
  • /v1/reviews/{id}/l1
  • /v1/reviews/{id}/assign-l2
  • /v1/reviews/{id}/l2
  • /v1/reviews/{id}/approve
  • /v1/reviews/{id}/merge

4.2 Release Lifecycle

States:

  • planned -> frozen -> rc -> validating -> released
  • Failure/rollback branches: failed, rolled_back

Main APIs:

  • /v1/releases/{id}/freeze
  • /v1/releases/{id}/rc
  • /v1/releases/{id}/validate
  • /v1/releases/{id}/publish
  • /v1/releases/{id}/rollback

4.3 Evaluation Lifecycle

Flow:

  1. Create run (/v1/evaluations/runs)
  2. Submit task results (/runs/{id}/results)
  3. Compute metrics + gate report
  4. Save baseline (/runs/{id}/baseline)
  5. Triage/resolve regressions

4.4 Autonomy Budget Lifecycle

States:

  • draft -> pending_approval -> active -> suspended|expired|completed

Flow:

  1. Create budget
  2. Activate or transition
  3. Run preflight checks
  4. Create enforcer
  5. Evaluate command/file/network requests
  6. Track escalations

4.5 Improvement Lifecycle

Research intake states:

  • submitted -> triaged -> analyzing -> accepted|deferred|rejected -> tracked

Associated loops:

  • Lesson capture and verification
  • Periodic checklist completion
  • Metrics snapshots and trend reporting
  • Roadmap refresh records

4.6 Trace Lifecycle

States:

  • running -> completed|failed|cancelled

Flow:

  1. Start trace
  2. Add step actions
  3. Record failures if present
  4. Complete trace

5. Known Integration Gaps

  • SessionSummarizer is exposed via session_summarizer_class in startup, but memory_search.summarize_session expects app.state.session_summarizer.
  • ObservationCapture is created without long-term memory/embedder bindings by default.
  • Tool registry/governance exists but is not fully attached as an app-wide execution path during startup.
  • Messaging adapters are not auto-registered in webhooks._adapters by default.
  • Training routes expect training_runner and agent_optimizer; default startup only guarantees training_store when enabled.
  • CLI agent33 run targets a legacy path (/api/v1/workflows/run) while workflow execution API is /v1/workflows/{name}/execute.

6. Pending and In-Progress Capabilities

The project is under active development, and some capabilities may be in progress on feature branches. This document reflects the current operational runtime on the active branch.