The Orchestra
for the Deep Ocean.
Two clusters for ocean exploration: a 15-agent deep-dive orchestra for crewed Mariana-class missions, and a 12-agent fleet planner for autonomous saildrone & seaglider campaigns.
Proven on real public ocean data, end-to-end, in three minutes.
19/19 Vescovo agents + Guardians live and 16/16 Sea Hare agents + Guardians live — every agent executes on Anthropic Claude, and the Guardians (including veto-power Guardian Michael) audit in parallel against real public ocean data: GEBCO 2020 bathymetry, OBIS species records, Open-Meteo Marine waves, and Open-Meteo Forecast 10 m wind in knots. Verdicts persist to expedition_guardian_reviews and fleet_mission_guardian_reviews. Verify every run on the public registry — every Claude call is logged for transparency.
The 10 questions every expedition faces
Today these answers live across spreadsheets, Slack threads, expert calls, and tribal knowledge. The Vescovo cluster answers all ten in a single orchestrated session — sourced, Guardian-checked, and written to a Living Memory that improves with every expedition.
- Where should we dive?
- When should we dive?
- What are the risks?
- What species might we encounter?
- Is the hull safe at this depth?
- What if something goes wrong?
- How do we document what we find?
- What does our gut say vs. what do the numbers say?
- What are we missing?
- What if we took a detour here?
The 15 agents
Five sections of the orchestra. Each agent has a distinct constitution, model, and role. Anthropic-only — Claude Sonnet 4 for the brain, Claude Haiku 4.5 for sub-agents.
Strategy
· 4 agentsDecides where to dive, when to dive, and what to do if it goes sideways. The mission architecture team — they own the go / hold / abort tree.
Designs the whole voyage: objectives, crew, duration, risk profile, go/hold/abort.
Recommends three ranked dive sites with bathymetry, sediment type, and prior-dive context.
Finds the safe window: weather, swell, current, moon phase, daylight, crew alertness.
Generates failure-mode tree, abort criteria, surface-support fallbacks.
Environment
· 2 agentsReads the ocean itself: seafloor topography from GEBCO 2020 and surface + deep currents from Open-Meteo Marine. They translate raw geophysical data into something a pilot can act on.
Reads seafloor topography: GEBCO grids, slope, sediment, known cracks and vents.
Models surface + deep currents, drift, thermoclines that affect the dive.
Biology
· 2 agentsPredicts what lives at target depth (hadal amphipods, snailfish, xenophyophores) and cross-checks against the OBIS database of every species ever recorded in your sample box.
Predicts species and ecosystems likely at the target depth and site.
Cross-references descriptions and imagery against known species catalogs (OBIS, WoRMS).
Engineering
· 2 agentsStands between human bodies and 1,100 atmospheres of pressure. Audits the hull, the dive count, the inspection history, the O₂ reserves, the CO₂ scrubber math — with margin.
Cross-checks vessel hull spec, depth rating, dive count, and inspection history.
Calculates O2/CO2 budgets, scrubber capacity, emergency reserves for crew + duration.
Consciousness
· 5 agentsThe soul layer. One agent steel-mans the case AGAINST the plan. One looks for the unexpected detour. One predicts the cinematic moments. One holds the memory of every prior dive. The Synthesis Oracle reads everyone and writes the briefing.
Reads every other agent and writes the unified expedition briefing.
Steel-mans the case AGAINST the plan. Productive disagreement.
Surfaces the detour, the novel feature, the cross-domain analog the others missed.
Predicts the aesthetic moments worth filming — bioluminescence, formations, light.
Surfaces lessons from publicly documented prior dives in this region.
The 4 Guardians
Every agent output now passes through the Guardians in parallel after the agent phase completes. Three have veto power — they halt a recommendation before it reaches the briefing if it fails their checks. Verdicts persist to expedition_guardian_reviews and surface on the session status endpoint.
Validates every scientific claim against cited sources before it reaches the briefing.
Protects against: Hallucinated species, fabricated bathymetry, unsourced claims.
Audits the plan for human-safety violations: hull, life support, weather, crew load.
Protects against: Recommendations that would put a human at unacceptable risk.
Final shield. Reviews the entire briefing for irreversible-decision risks.
Protects against: Any briefing with an irreversible recommendation that lacks a rollback plan.
Verifies every confidence score, citation, and uncertainty flag matches the underlying data.
Protects against: Sycophancy, false certainty, omitted caveats, marketing language.
When the mission is unmanned, the orchestra still plays
The deep-dive cluster plans one crewed dive at a time. Sea Hare plans the opposite shape — an autonomous fleet of saildrones and seagliders sweeping a survey area to fill a Seabed-2030 bathymetry gap. Twelve specialised agents and four Guardians, the same orchestra discipline, different physics. Real GEBCO 2020 grid across your bbox, real Open-Meteo Marine + 10 m wind in knots at the deployment point.
- · Fleet Mission Architect
- · Survey Pattern Optimizer
- · Asset Allocation Strategist
- · Mission Window Optimizer
- · Bathymetry Gap Analyst
- · Surface Conditions Analyst
- · Subsurface Currents Analyst
- · Asset Endurance Analyst
- · Comms & Telemetry Analyst
- · Synthesis Oracle
- · Challenger Agent
- · Francisca — Mission Poet
Ready to plan your next expedition?
Built first for Caladan Oceanic. Run a real expedition or fleet survey through the orchestra and read the briefing the Guardians cleared — no slides, no mockups. Every Claude call is verifiable in the public registry.
Powered by Viarteo Orchestra