Skip to content

README.md

Latest commit

 

History

History
138 lines (103 loc) · 3.61 KB

README.md

File metadata and controls

138 lines (103 loc) · 3.61 KB
Raw

Evolution Example

Demonstrates the Evolution (Agent Self-Learning) feature of trpc-agent-go with a real LLM and live skill extraction.

What It Does

  1. Runs multiple rounds of similar tasks (city info lookup)
  2. After each task, the background reviewer extracts a reusable skill
  3. On subsequent rounds, the agent loads the skill via skill_load
  4. You can observe: skill creation, gate metrics, and warm-start behavior

Running

export OPENAI_API_KEY="sk-..."
cd examples/evolution
go run main.go

Options:

go run main.go -model gpt-4o-mini -rounds 5
go run main.go -clean  # reset skills and start fresh

Expected Output

============================================================
Round 1/3
Task: Look up the population and country of Tokyo, then summarize...
Skills available: 0
------------------------------------------------------------

Response: Tokyo has a population of 13.96 million and is in Japan.
Tool calls: 1 | Time: 2.3s
Waiting for background skill extraction... done.

============================================================
Round 2/3
Task: Look up the population and country of Paris, then summarize...
Skills available: 1 [City Info Lookup]
------------------------------------------------------------

Response: Paris has a population of 2.16 million and is in France.
Tool calls: 1 | Time: 1.8s
Waiting for background skill extraction... done.

...

============================================================
FINAL STATE
============================================================

Managed skills (1):
  - City Info Lookup: Looks up city facts using city_lookup tool...

Quality gate metrics:
  Candidates seen:       2
  Revisions promoted:    2
  Spec-gate rejected:    0
  Safety-gate rejected:  0

Re-run the same example to see warm-start behavior (skills loaded immediately).

Key Observations

  • Round 1: no skills → agent figures it out from scratch
  • Round 2+: skill loaded → agent has a "checklist" to follow
  • Re-run: warm-start, skill available from the first task
  • Gate metrics: shows how many candidates passed quality checks

Configuration Variants

Add Human Approval

evolution.WithHumanGate(evolution.NewCreateOnlyHoldGate())

New skills will be held in pending_approval state until you approve them programmatically:

svc := evolution.NewApprovalService(store, pointer, nil)
pending, _ := svc.ListPending(ctx, evolution.ListPendingOpts{})
svc.Decide(ctx, evolution.ApprovalDecision{
    RevisionID: pending[0].RevisionID,
    SkillID:    pending[0].SkillID,
    Approved:   true,
    Reviewer:   "you@example.com",
})

Customize Review Trigger

The default review policy runs the reviewer for richer session deltas: at least 4 tool calls, user correction, or recovered tool error. You can tune the built-in threshold:

evolution.WithReviewPolicy(evolution.DefaultReviewPolicy{
    MinToolCalls: 6,
})

For centralized routing, implement evolution.ReviewPolicy and use the context.Context plus evolution.ReviewPolicyInput metadata to call your decision service.

Shadow Mode (observe without enforcing)

evolution.WithApprovalGateShadow(true)

Gates evaluate and log but never block promotion — useful for rollout.

Files Generated

After running, you'll see:

managed_skills/
  city-info-lookup/
    SKILL.md              ← the extracted skill

managed_skills_revisions/
  city-info-lookup/
    revisions/
      20260501T.../
        meta.json         ← immutable revision snapshot
    audit.log             ← append-only audit trail
    active.txt            ← points to current revision