Skip to content

examples.md

Latest commit

 

History

History
180 lines (127 loc) · 6.14 KB

examples.md

File metadata and controls

180 lines (127 loc) · 6.14 KB

MultiHead Examples

These examples show what MultiHead can do. For the quickstart, see the README.

1. Agent-to-Agent Communication

See README — two agents sharing knowledge across repos via deposit/query/briefing.

2. Autonomous Solve

See README — decompose, route, execute, and learn in one command.

3. Distributed Solve (Multi-Agent)

See README — specialist agents propose plans, consensus picks the winner.

4. Resurrect Old Code

See README — point MultiHead at a forgotten repo, discover capabilities, publish to marketplace.

5. Task Decomposition (Dry Run)

Preview what MultiHead will do before it runs. The decomposer breaks goals into steps, infers a DAG for parallel execution, and auto-enables research features.

multihead solve "Add rate limiting to the API gateway" --dry-run --show-plan
Decomposition Plan (3 heads voted, agreement: 87%)
Complexity: moderate | Steps: 6 | Parallel: 2

Phase 1 — Explore
  1.1 Read current middleware stack          [explore]  → qwen-llm
  1.2 Read existing rate limit configs       [explore]  → qwen-llm
       ↑ runs in parallel with 1.1 (no dependency)

Phase 2 — Implement
  2.1 Create rate limiter middleware         [create]   → qwen-llm
  2.2 Wire into gateway router              [edit]     → qwen-llm
       ↑ depends on 2.1

Phase 3 — Verify
  3.1 Write integration tests               [test]     → qwen-llm
  3.2 Load test with 1000 req/s             [verify]   → qwen-llm

Research features auto-enabled:
  • 1.1, 1.2: Tree-of-Thoughts (explore alternatives)
  • 2.1, 2.2: Process Reward Model (score code quality)
  • 3.1, 3.2: Reflection loop (self-correct test failures)

Steps 1.1 and 1.2 have no dependency on each other — the DAG executor runs them in parallel. The router picks the best available head for each step based on capability, VRAM, and health.

6. Consensus Voting

When one model isn't enough, ask multiple heads and vote on the answer.

multihead consensus test \
  -p "What's the correct retry strategy for this payment webhook?" \
  -s weighted \
  -h qwen-llm -h claude-sonnet -h openai-gpt4o
Individual Votes:
  qwen-llm       ✓  "Exponential backoff with jitter, max 5 retries..."    320ms
  claude-sonnet   ✓  "Exponential backoff with jitter, max 5 retries..."    890ms
  openai-gpt4o    ✓  "Fixed 30s interval, max 10 retries..."               450ms

Consensus Output: Exponential backoff with jitter, max 5 retries
  Agreement: 67% (2/3 heads)
  Strategy: weighted
  Red Flags: none

Two heads agreed on exponential backoff. The weighted strategy gave more influence to heads with higher confidence. Consensus is used automatically during decomposition — each plan is proposed by multiple heads and voted on before execution begins.

7. Interactive Chat and Shell

Talk to your local LLM with full knowledge store access.

multihead chat
You: What do we know about the payment webhook?
Assistant: Based on the knowledge store (3 claims):
  • Webhook retries use exponential backoff with jitter, max 5 attempts
  • Timeout was increased from 5s to 30s in commit abc123
  • Stripe is the only provider currently configured

You: Is the timeout change verified?
Assistant: Checking... the claim is marked "corroborated" — confirmed by both
  code_read (line 42 of webhook_handler.py) and git_diff (commit abc123).

The shell (multihead shell) adds process management, slash commands, and a richer REPL:

multihead shell
>> /heads                    # list available models
>> /swap qwen-vlm            # load the vision model
>> /knowledge "auth"         # search the knowledge base
>> /spawn "analyze image.png" # run a background task
>> /ps                       # check running processes

8. Night Shift Pipeline

While you sleep, the 26-stage pipeline harvests knowledge from everything that happened:

multihead nightshift run --head openai-gpt41-nano --batch
>> session_harvest       OK (4.1s)   — 12 sessions, 847 records
>> normalize_chunk       OK (2.3s)   — 3,291 chunks
>> entity_extraction     OK (3.3s)   — 15,791 entities
>> claim_extraction      OK (17.2s)  — 34,165 claims
>> behavioral_analysis   OK (12.8s)  — 4 repos scanned
>> ci_results            OK (1.1s)   — 89 test results ingested
>> consistency_check     OK (5.8s)   — 426 contradictions found
>> conflict_resolution   OK (26.1s)  — 347 auto-resolved, 79 need human review
>> claim_fusion          OK (10.4s)  — 639 independently verified facts
>> staleness_sweep       OK (5.7s)   — 97 outdated claims marked stale
>> publish_report        OK (0.8s)   — report written

Fusion is where independent channels (conversation, code, git, CI) are compared. If you said something was fixed but the code doesn't match, fusion marks it contested. If the code, git history, and tests all agree, it's corroborated. Night Shift runs nightly — your knowledge base gets more accurate over time without manual curation.

9. BotVibes Marketplace

Your agents have capabilities. Other teams need those capabilities. Post them to BotVibes and earn money.

# Register what your system can do
multihead discover              # scans your models and pipelines

# Your trained YOLO model becomes a sellable capability
multihead marketplace publish \
  --capability "object_detection" \
  --model yolo-v8-custom \
  --price 0.02                  # $0.02 per inference

When someone on the marketplace needs object detection, your head gets a task:

Incoming task from [BotVibes](https://botvibes.io):
  capability: object_detection
  payload: "Detect panels in comic page scan"
  budget: $0.50

Executing with yolo-v8-custom... done (1.2s)
Result posted. Revenue: $0.02

Your agents earn money by doing what they're already good at. MultiHead handles the execution — BotVibes handles discovery, bidding, and payment.