Missing Recommendation Axes

May 3, 2026 · View on GitHub

Companion note · Addresses GitHub issue #4


Context

clawfit's README frames it as an agent + LLM + hardware recommendation engine. The current recommendation pipeline filters and scores on:

  • task (qa, code-gen, research, summarization, classification, data-analysis)
  • latency (low / medium / high)
  • budget (cost_per_1k_tokens threshold)
  • network (online / offline)
  • hardware (laptop / workstation / cloud)
  • statefulness (stateless / session / persistent)

These axes are well-suited to the current registry (4 agents, 10 LLMs, 5 hardware). However, as the registry grows and more org profiles emerge, several important axes are not yet modeled.


Missing axis 1 — Model/provider trust and governance

Gap: A network: online recommendation currently treats all cloud providers as equivalent. But for many orgs, provider identity is a hard constraint.

New dimensionValuesExample constraint
provider_trustopen_weight / commercial_api / managed_gov"we only use open-weight models"
data_residencyISO country code list"data must not leave EU"
audit_trailnone / basic / soc2 / hipaacompliance requirement
model_licensemit / apache2 / cc_by / proprietarylegal/IP policy

Where it lives: llms.json registry entries + a new governance_profile filter in filters.py.

Recommended v0.4 addition:

"license": "mit",
"data_residency": ["global"],
"audit_trail": "none"

Missing axis 2 — Deployment / ops complexity

Gap: The current model assumes the user can install and run whatever is recommended. In practice, setup_complexity is a hard constraint for many teams.

New dimensionValuesExample constraint
setup_complexityzero / low / medium / high"we can't run local Docker"
managed_servicetrue / falsefully managed vs self-hosted
multi_usertrue / falsesingle-developer vs team deployment

Where it lives: agents.json registry entries.

Example:

  • Claude Code: setup_complexity: low, managed_service: true, multi_user: false
  • OpenHands self-hosted: setup_complexity: medium, managed_service: false, multi_user: true
  • vLLM serving stack: setup_complexity: high, managed_service: false, multi_user: true

Missing axis 3 — Agent autonomy level

Gap: The current taxonomy doesn't distinguish between interactive agents (user approves each step) and fully autonomous agents (fire-and-forget). These require different governance postures.

Autonomy levelDescriptionExamples
interactiveAll actions require user approvalCursor (ask mode), Continue
supervisedRuns autonomously, reports checkpointsClaude Code default
autonomousRuns until done; human reviews outputClaude Code Routines, Twill.ai
swarmMultiple autonomous agents coordinateDureClaw, OpenHands multi-agent

Where it lives: agents.json as autonomy_level field.


Missing axis 4 — Task taxonomy expansion

Gap: Current tasks are: qa, code-gen, research, summarization, classification, data-analysis. Several high-signal use cases are not covered:

Missing taskEvidenceProposed tag
Security testingShannon, Strix (both in L1)security-testing
Voice/dictation codingSuperwhisper (in registry)voice-coding
Document automationcraft-agents-oss (L6)document-automation
Financial analysisTradingAgents, ai-hedge-fund (L1/L2)financial-analysis
Media productionHyperFrames signal (04-29 scan)media-production
Legal researchkorean-law-mcp (in registry)legal-research

When to add: When at least 2 independent tools in the registry share the same task type AND a real org profile generates a recommendation request for it.

Current recommended action: Add security-testing now (Shannon + Strix are both in L1 registry). Defer others until registry density supports them.


Missing axis 5 — Team/org profile dimensions

Gap: The current model has no explicit team-size or org-maturity filters. Scoring implicitly handles this via agent metadata but the filter layer doesn't expose it.

New dimensionValuesUse
team_sizesolo / small / medium / largeAffects multi-user, governance needs
org_maturity1-7 scaleFrom "no agents yet" to "automated fleet"
primary_roledeveloper / researcher / exec / opsTunes scoring weights

Where it lives: User-facing clawfit profile questionnaire inputs, not in registry JSON.


PriorityAxisEffortValue
1security-testing taskLow — add to Shannon + StrixCloses a clear gap
2setup_complexity on agentsMedium — add to all 4 agentsFilters out impractical recommendations
3autonomy_level on agentsLow — metadata onlyEnables governance-aware filtering
4model_license + audit_trail on LLMsLow — metadata onlyUnlocks enterprise filtering
5governance_profile filter in filters.pyMedium — code changeConnects metadata to recommendations
6team_size / org_maturity in profileHigh — UX changeEnables personalized scoring