Build the tender picture.
Agents gather tender requirements, vessel availability, vetting and crew records, and surface missing evidence for the team to resolve.
CoralOS
Many agents on one problem, each with a clear job.
AI has scaled vertically: bigger models, longer context, more reasoning per call. Coral works on the horizontal scaling axis. Agents of different kinds, often built by different teams, work one problem with clear responsibility for each part. The gain comes from how they are organised, not how many there are.
Drag the slider. Switch the organisation.
What Coral includes
One server underneath. Applications on top.
Any modelAny frameworkAny harness
Click a layer to see what it does.
Coral Discovery · Battery experiment
Better battery rules, measured in replay.
An agent team improved the controller that bids and dispatches a solar, wind and battery plant at Khavda, Gujarat. The test replays 271 days of 2025 with real market prices and real weather on a modelled plant.
+13.5%
modelled net income over the starting rule
$156.7M
over 271 days, against $138.0M for the starting rule
MW is the rate of power; MWh is stored energy. Battery size follows the announced Khavda project. Plant output comes from a physical model driven by site weather.
The decision
Each day the controller makes a promise, then steers the battery to keep it.
Day before, 10:00
Promise
Commit tomorrow's export for all 96 quarter-hour blocks, before tomorrow's weather or prices are known.
Every 15 minutes
Steer
Charge, discharge or wait, within the battery's power and stored energy, to close the gap between the promise and what the plant produces.
After delivery
Settle
Earn the day-ahead price for the promise. Pay or receive settlement for deviations. Count the wear each cycle puts on the battery.
Net income = day-ahead income + deviation settlement − battery wear
The result
Better rules earned more from the same plant on the same 271 days.
Net income, INR crore, 271 test days
Bars start at zero. One crore is INR 10 million.
+158.6
crore more than the starting rule, or +13.5%
Where the money came from
The best controller earned slightly less day-ahead income than earlier versions. Smaller settlement costs and less battery wear more than made up for it.
| Controller version | Day-ahead | Settlement | Wear | Net |
|---|
INR crore. The starting rule and the single agent's records have no breakdown.
Replay the search
The team went back to an approach the single agent had dropped, and it paid off.
The single agent tried a battery optimiser once, saw it fail, and never returned. Its best came at attempt 45. The four-agent team opened nine directions and got the battery optimiser working after 16 attempts that didn't beat the best.
Key changes
Best so far
One agent's best, 1,305.9
Agent 1Agent 2Agent 3Agent 4
▾ below 1,150 · × failed
4 agents9 directions103 attempts, 100 scoredOne agent: 54 attempts, best at 45
Replay a day
Play the same day through an early and the best controller.
Both versions see the same weather and prices. Watch how each one charges, discharges and tracks its promise through the day.
Day-ahead price
Solar + wind
Promised export
Actual export
Battery charge level
Replayed experiment output, not live plant data. Day tallies are before battery wear, which is only scored over the full period.
How the test works
Only the controller changes. A fixed, private grader scores every attempt the same way.
What is real
- IEX day-ahead, real-time and green day-ahead prices for every quarter hour, 2023 to 28 September 2025
- ERA5 weather at the Khavda site, with the previous day's forecasts and their errors
- Battery size of the announced 1,126 MW / 3,530 MWh Khavda project
- Training on 2023–2024. Scoring on 1 January to 28 September 2025
What is modelled or simplified
- Plant output comes from a physical model of tracker PV and a wind power curve, not the real plant's meters
- Settlement uses simplified bands from the 2024 deviation rules
- Wear is a proxy: cycles weighted by depth, battery at $125/kWh over 6,000 cycles
- Bids clear in full at market price. No transmission charges, fees, taxes or failures
- The controller runs sandboxed, with no network and no access to outside files
Sources
$ figures convert INR at 85 per USD. The $25M a year straight-line annualises the gain over the starting rule; it is not booked revenue.
The decision changes. The way you test a better rule stays the same.
A battery controller schedules energy under physical limits. A vessel planner schedules vessels under operating limits. In both cases, Coral Discovery searches for a better rule against your team's own test.
One round of the search
Each agent works in rounds, and every round leaves a scored, saved attempt.
Team size is a budget setting
Who does what
Your engineers own the decision and the test. Coral brings the search.
You bring
- One recurring decision worth improving
- A test you trust: hard limits and one number
- Today's rule as the starting point
- Your infrastructure and your choice of models
Coral brings
- The agent team and the shared board
- Engineers who support your team as it runs the search
You keep
- The improved rule
- Its pass against your test
- The score history
- Every direction tried, as a replayable record
Same method, different decisions
Each decision needs its own data, limits and test.
The battery and vessel results are measured experiment output on modelled plants and markets. The crew page is an illustration on small example data. A score in one domain says nothing about savings in another.
For Astro Offshore
Better tender decisions.
Better tender decisions.
A fleet that can deliver.
A tender win is only valuable if you can deliver it profitably. Explore how bidding, vessel commitments and crew availability change the return from the same opportunities.
A recorded, interactive experiment. Public Astro vessel names and specifications; simulated opportunities, costs and events.
Bid with the whole fleet in view.
12named vessels
40modelled tenders
90simulated days
See the trade-off between winning work, keeping vessels available and paying for external cover.
This is a public-fleet-inspired showcase, not a replay of Astro’s operating history.
Replay 90 days
Play the same 90 days. Compare the decisions.
Both approaches face the same opportunities and disruptions. Start with the fleet, then inspect tendering or crew availability. One playback control keeps the two approaches on the same day.
The fleet in this scenario
DP is the dynamic positioning class: 0 = none (anchored or moored), DP1 = one position-keeping system, DP2 = redundant systems (required alongside platforms), DP3 = redundant and physically separated. Bollard pull (t) is towing force; deck in m²; crane in tonnes; berths = people the vessel can accommodate.
Use Play, pause on a decision, or drag the day marker. Actual vessel activity and outcomes appear through the selected day; tender plans appear when announced. On narrow screens, scroll each timeline sideways.
Recorded simulator outputs. “Research winner” means the best average result on the research benchmark; another candidate can perform better on an individual scenario. Figures on this panel belong to the selected scenario and are rounded. They are not Astro operating results.
How the rule was foundWatch the Discovery experiment46 recorded attempts · four agents · optional research detail
Replay the search
Watch Discovery find a better rule.
Four agents tested 46 changes across 13 lines of enquiry. Play the recorded search to see the gains, the unsuccessful changes and the point where a new bidding approach improved the benchmark result. This is a replay of completed research, not a live agent run.
Key changes
Best so far
Agent 1Agent 2Agent 3Agent 4
New best at that attempt
4 agents13 directions46 attempts, 46 scored2 h 18 min, best at attempt 40
The result
The research benchmark: +15.0% over the starting rule.
Reported mean modelled contribution, MUSD per scenario, 12 research scenarios
The 12 private-seed cases were used to score candidates during the search. They were hidden from the agents, but provided repeated score feedback; this is not an untouched final holdout. These averages are separate from the Astro showcase above.
+0.61
MUSD more per scenario than the starting rule, or +15.0%
Where the money came from
The first gains came from dropping blanket reservations and bridging crew gaps with agency crew, which cut external cover. The best version then bid on far more tenders: charter revenue rose by a tenth, and it paid more cover for the contracts it could not serve itself.
| Policy version | Revenue | Operating | Cover | Crew & vetting | Net |
|---|
Reported rounded components, MUSD per scenario over 12 research cases. Component totals are approximate and should not be treated as an exact reconciliation of the reported score. Operating includes opex and transit fuel.
Beyond the experimentDiscovery improves the rule.
Discovery improves the rule.
CoralOS connects the work.
The same tender decision crosses commercial, fleet and crewing teams. CoralOS can coordinate agents across that workflow, with people approving bids and commitments. These are proposed workflows for an Astro pilot.
Compare the choices.
Discovery tests changes to the bidding and assignment rule against agreed limits. Decision-makers inspect the trade-offs and approve the approach.
Keep the decision connected.
Coordinated agents can prepare the bid pack, update the proposed fleet and crew plan, and flag changes that require human approval.
Illustrative workflow. No live connection to Astro’s systems; no bids, staffing changes or commitments are made by this demo.
Understand the assumptions and evidence
The decision
Each morning the policy bids, assigns and keeps its promises with the fleet it has.
Every morning
Bid
Pick open tenders and the vessel to offer. Reserve the vessel's window, start client vetting in time, or withdraw a bid that no longer fits.
On award
Assign
Put a compliant vessel on the contract so it reaches the field inside the delivery window, or buy external cover at 1.35 times the contracted day rate plus a mobilisation charge.
Through the day
Keep the promise
Crews rotate every 28 days and leave can create gaps; weather closes fields; breakdowns void assignments. Hire agency crew, re-assign, or cover.
Contribution = charter revenue − operating cost − transit fuel − external cover − damages − agency crew − vetting
Research score = mean across 12 private-seed scenarios of (contribution in MUSD − 5 × feasibility violations)
How the test works
What this demonstration can tell you.
What is real
- Vessel classes and the capability thresholds of each job type (DP class, bollard pull, deck area, crane, berths) from Astro Offshore's fleet pages and spec sheets and from ONGC and Indian port charter tenders
- Day-rate ranges from Tidewater's SEC filings, Seabrokers' North Sea report and Clarksons / M3 Marine surveys; the tender time chain from public Indian tenders; contract terms (mobilisation, cover, options) from BIMCO SUPPLYTIME 2017
- Client vetting from OCIMF OVID and IMCA eCMID practice; 28-day crew rotations from the Strathclyde vessel crew scheduling data
- The Astro showcase uses 12 public vessel names and a supplied specification table. Starting locations and availability are showcase assumptions; the supplied position snapshot needs a dated source and reconciliation before being treated as operational evidence.
What is modelled or simplified
- Tenders, clients, awards, delays, weather, breakdowns, crew leave and costs are modelled. The Astro showcase is not Astro’s actual tender pipeline, crew roster, financial return or 90-day operating history.
- Operating cost and transit fuel per vessel-day are set by us; prices do not react to bids and there is no price-to-win model (a bid is won or lost by a hidden draw)
- Transit time is great-circle distance times 1.15 at service speed; the Astro showcase drops rig moves and crew transfer because the real 12 include no 150 t AHTS and no crew boat
- The research describes sandboxed candidate policies with no network access. A future Astro pilot needs an agreed current rule, validated records and a separate final evaluation period.
Sources
Results and replay records were supplied with this demonstration. MUSD means millions of US dollars of modelled contribution, not booked revenue. The winning policy and full research run have not been independently reproduced here. The research was stopped after 46 scored attempts.
An Astro pilot would use your own vessel, charter and crew records, compare against your current rule and agree an independent final evaluation period with your engineers.
Back to the beginning ↑
Coral Discovery · Crew rotation
Qualified people, feasible rotations.
A covered charter also needs a crew that holds the right certificates, DP experience and visas, and has had its rest. Coverage comes first; only then does the cheaper rotation win.
Illustrative example on four made-up crew groups with declared example costs. These are not Astro personnel or pay rates.
Charter
A cheaper rotation that breaks a rule is not a result.
The current rule picks the first crew that passes every check. The alternative picks the lowest-cost crew that passes every check.
Costs are declared example figures per rotation: travel + relief + overtime. Certificates, visas, availability and rest are hard limits. One crew group covers the required roles in this simplified example.